blob: 337c6afb3345e1e0d6071a63718f4dc6aac140b7 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Uwe Zeisbergerf30c2262006-10-03 23:01:26 +02002 * mm/page-writeback.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2002, Linus Torvalds.
Peter Zijlstra90eec102015-11-16 11:08:45 +01005 * Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
Linus Torvalds1da177e2005-04-16 15:20:36 -07006 *
7 * Contains functions related to writing back dirty pages at the
8 * address_space level.
9 *
Francois Camie1f8e872008-10-15 22:01:59 -070010 * 10Apr2002 Andrew Morton
Linus Torvalds1da177e2005-04-16 15:20:36 -070011 * Initial version
12 */
13
14#include <linux/kernel.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040015#include <linux/export.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070016#include <linux/spinlock.h>
17#include <linux/fs.h>
18#include <linux/mm.h>
19#include <linux/swap.h>
20#include <linux/slab.h>
21#include <linux/pagemap.h>
22#include <linux/writeback.h>
23#include <linux/init.h>
24#include <linux/backing-dev.h>
Andrew Morton55e829a2006-12-10 02:19:27 -080025#include <linux/task_io_accounting_ops.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <linux/blkdev.h>
27#include <linux/mpage.h>
Peter Zijlstrad08b3852006-09-25 23:30:57 -070028#include <linux/rmap.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029#include <linux/percpu.h>
30#include <linux/notifier.h>
31#include <linux/smp.h>
32#include <linux/sysctl.h>
33#include <linux/cpu.h>
34#include <linux/syscalls.h>
Al Viroff01bb42011-09-16 02:31:11 -040035#include <linux/buffer_head.h> /* __set_page_dirty_buffers */
David Howells811d7362006-08-29 19:06:09 +010036#include <linux/pagevec.h>
Jan Karaeb608e32012-05-24 18:59:11 +020037#include <linux/timer.h>
Clark Williams8bd75c72013-02-07 09:47:07 -060038#include <linux/sched/rt.h>
Ingo Molnarf361bf42017-02-03 23:47:37 +010039#include <linux/sched/signal.h>
Lisa Du6e543d52013-09-11 14:22:36 -070040#include <linux/mm_inline.h>
Dave Chinner028c2dd2010-07-07 13:24:07 +100041#include <trace/events/writeback.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042
Lisa Du6e543d52013-09-11 14:22:36 -070043#include "internal.h"
44
Linus Torvalds1da177e2005-04-16 15:20:36 -070045/*
Wu Fengguangffd1f602011-06-19 22:18:42 -060046 * Sleep at most 200ms at a time in balance_dirty_pages().
47 */
48#define MAX_PAUSE max(HZ/5, 1)
49
50/*
Wu Fengguang5b9b3572011-12-06 13:17:17 -060051 * Try to keep balance_dirty_pages() call intervals higher than this many pages
52 * by raising pause time to max_pause when falls below it.
53 */
54#define DIRTY_POLL_THRESH (128 >> (PAGE_SHIFT - 10))
55
56/*
Wu Fengguange98be2d2010-08-29 11:22:30 -060057 * Estimate write bandwidth at 200ms intervals.
58 */
59#define BANDWIDTH_INTERVAL max(HZ/5, 1)
60
Wu Fengguang6c14ae12011-03-02 16:04:18 -060061#define RATELIMIT_CALC_SHIFT 10
62
Wu Fengguange98be2d2010-08-29 11:22:30 -060063/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070064 * After a CPU has dirtied this many pages, balance_dirty_pages_ratelimited
65 * will look to see if it needs to force writeback or throttling.
66 */
67static long ratelimit_pages = 32;
68
Linus Torvalds1da177e2005-04-16 15:20:36 -070069/* The following parameters are exported via /proc/sys/vm */
70
71/*
Jens Axboe5b0830c2009-09-23 19:37:09 +020072 * Start background writeback (via writeback threads) at this percentage
Linus Torvalds1da177e2005-04-16 15:20:36 -070073 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080074int dirty_background_ratio = 10;
Linus Torvalds1da177e2005-04-16 15:20:36 -070075
76/*
David Rientjes2da02992009-01-06 14:39:31 -080077 * dirty_background_bytes starts at 0 (disabled) so that it is a function of
78 * dirty_background_ratio * the amount of dirtyable memory
79 */
80unsigned long dirty_background_bytes;
81
82/*
Bron Gondwana195cf4532008-02-04 22:29:20 -080083 * free highmem will not be subtracted from the total free memory
84 * for calculating free ratios if vm_highmem_is_dirtyable is true
85 */
86int vm_highmem_is_dirtyable;
87
88/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070089 * The generator of dirty data starts writeback at this percentage
90 */
Wu Fengguang1b5e62b2009-03-23 08:57:38 +080091int vm_dirty_ratio = 20;
Linus Torvalds1da177e2005-04-16 15:20:36 -070092
93/*
David Rientjes2da02992009-01-06 14:39:31 -080094 * vm_dirty_bytes starts at 0 (disabled) so that it is a function of
95 * vm_dirty_ratio * the amount of dirtyable memory
96 */
97unsigned long vm_dirty_bytes;
98
99/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700100 * The interval between `kupdate'-style writebacks
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700102unsigned int dirty_writeback_interval = 5 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103
Artem Bityutskiy91913a22012-03-21 22:33:00 -0400104EXPORT_SYMBOL_GPL(dirty_writeback_interval);
105
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106/*
Alexey Dobriyan704503d2009-03-31 15:23:18 -0700107 * The longest time for which data is allowed to remain dirty
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108 */
Toshiyuki Okajima22ef37e2009-05-16 22:56:28 -0700109unsigned int dirty_expire_interval = 30 * 100; /* centiseconds */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110
111/*
112 * Flag that makes the machine dump writes/reads and block dirtyings.
113 */
114int block_dump;
115
116/*
Bart Samweled5b43f2006-03-24 03:15:49 -0800117 * Flag that puts the machine in "laptop mode". Doubles as a timeout in jiffies:
118 * a full sync is triggered after this time elapses without any disk activity.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119 */
120int laptop_mode;
121
122EXPORT_SYMBOL(laptop_mode);
123
124/* End of sysctl-exported parameters */
125
Tejun Heodcc25ae2015-05-22 18:23:22 -0400126struct wb_domain global_wb_domain;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700127
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400128/* consolidated parameters for balance_dirty_pages() and its subroutines */
129struct dirty_throttle_control {
Tejun Heoe9f07df2015-05-22 18:23:28 -0400130#ifdef CONFIG_CGROUP_WRITEBACK
131 struct wb_domain *dom;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400132 struct dirty_throttle_control *gdtc; /* only set in memcg dtc's */
Tejun Heoe9f07df2015-05-22 18:23:28 -0400133#endif
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400134 struct bdi_writeback *wb;
Tejun Heoe9770b32015-05-22 18:23:27 -0400135 struct fprop_local_percpu *wb_completions;
Jan Karaeb608e32012-05-24 18:59:11 +0200136
Tejun Heo9fc3a432015-05-22 18:23:30 -0400137 unsigned long avail; /* dirtyable */
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400138 unsigned long dirty; /* file_dirty + write + nfs */
139 unsigned long thresh; /* dirty threshold */
140 unsigned long bg_thresh; /* dirty background threshold */
141
142 unsigned long wb_dirty; /* per-wb counterparts */
143 unsigned long wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -0400144 unsigned long wb_bg_thresh;
Tejun Heodaddfa32015-05-22 18:23:26 -0400145
146 unsigned long pos_ratio;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400147};
148
Jan Karaeb608e32012-05-24 18:59:11 +0200149/*
150 * Length of period for aging writeout fractions of bdis. This is an
151 * arbitrarily chosen number. The longer the period, the slower fractions will
152 * reflect changes in current writeout rate.
153 */
154#define VM_COMPLETIONS_PERIOD_LEN (3*HZ)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700155
Tejun Heo693108a2015-05-22 17:13:49 -0400156#ifdef CONFIG_CGROUP_WRITEBACK
157
Tejun Heod60d1bd2015-09-29 12:47:53 -0400158#define GDTC_INIT(__wb) .wb = (__wb), \
159 .dom = &global_wb_domain, \
160 .wb_completions = &(__wb)->completions
161
Tejun Heo9fc3a432015-05-22 18:23:30 -0400162#define GDTC_INIT_NO_WB .dom = &global_wb_domain
Tejun Heod60d1bd2015-09-29 12:47:53 -0400163
164#define MDTC_INIT(__wb, __gdtc) .wb = (__wb), \
165 .dom = mem_cgroup_wb_domain(__wb), \
166 .wb_completions = &(__wb)->memcg_completions, \
167 .gdtc = __gdtc
Tejun Heoc2aa7232015-05-22 18:23:35 -0400168
169static bool mdtc_valid(struct dirty_throttle_control *dtc)
170{
171 return dtc->dom;
172}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400173
174static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
175{
176 return dtc->dom;
177}
178
Tejun Heo9fc3a432015-05-22 18:23:30 -0400179static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
180{
181 return mdtc->gdtc;
182}
183
Tejun Heo841710a2015-05-22 18:23:33 -0400184static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
185{
186 return &wb->memcg_completions;
187}
188
Tejun Heo693108a2015-05-22 17:13:49 -0400189static void wb_min_max_ratio(struct bdi_writeback *wb,
190 unsigned long *minp, unsigned long *maxp)
191{
192 unsigned long this_bw = wb->avg_write_bandwidth;
193 unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth);
194 unsigned long long min = wb->bdi->min_ratio;
195 unsigned long long max = wb->bdi->max_ratio;
196
197 /*
198 * @wb may already be clean by the time control reaches here and
199 * the total may not include its bw.
200 */
201 if (this_bw < tot_bw) {
202 if (min) {
203 min *= this_bw;
204 do_div(min, tot_bw);
205 }
206 if (max < 100) {
207 max *= this_bw;
208 do_div(max, tot_bw);
209 }
210 }
211
212 *minp = min;
213 *maxp = max;
214}
215
216#else /* CONFIG_CGROUP_WRITEBACK */
217
Tejun Heod60d1bd2015-09-29 12:47:53 -0400218#define GDTC_INIT(__wb) .wb = (__wb), \
219 .wb_completions = &(__wb)->completions
Tejun Heo9fc3a432015-05-22 18:23:30 -0400220#define GDTC_INIT_NO_WB
Tejun Heoc2aa7232015-05-22 18:23:35 -0400221#define MDTC_INIT(__wb, __gdtc)
222
223static bool mdtc_valid(struct dirty_throttle_control *dtc)
224{
225 return false;
226}
Tejun Heoe9f07df2015-05-22 18:23:28 -0400227
228static struct wb_domain *dtc_dom(struct dirty_throttle_control *dtc)
229{
230 return &global_wb_domain;
231}
232
Tejun Heo9fc3a432015-05-22 18:23:30 -0400233static struct dirty_throttle_control *mdtc_gdtc(struct dirty_throttle_control *mdtc)
234{
235 return NULL;
236}
237
Tejun Heo841710a2015-05-22 18:23:33 -0400238static struct fprop_local_percpu *wb_memcg_completions(struct bdi_writeback *wb)
239{
240 return NULL;
241}
242
Tejun Heo693108a2015-05-22 17:13:49 -0400243static void wb_min_max_ratio(struct bdi_writeback *wb,
244 unsigned long *minp, unsigned long *maxp)
245{
246 *minp = wb->bdi->min_ratio;
247 *maxp = wb->bdi->max_ratio;
248}
249
250#endif /* CONFIG_CGROUP_WRITEBACK */
251
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700252/*
Johannes Weinera756cf52012-01-10 15:07:49 -0800253 * In a memory zone, there is a certain amount of pages we consider
254 * available for the page cache, which is essentially the number of
255 * free and reclaimable pages, minus some zone reserves to protect
256 * lowmem and the ability to uphold the zone's watermarks without
257 * requiring writeback.
258 *
259 * This number of dirtyable pages is the base value of which the
260 * user-configurable dirty ratio is the effictive number of pages that
261 * are allowed to be actually dirtied. Per individual zone, or
262 * globally by using the sum of dirtyable pages over all zones.
263 *
264 * Because the user is allowed to specify the dirty limit globally as
265 * absolute number of bytes, calculating the per-zone dirty limit can
266 * require translating the configured limit into a percentage of
267 * global dirtyable memory first.
268 */
269
Johannes Weinera8045522014-01-29 14:05:39 -0800270/**
Mel Gorman281e3722016-07-28 15:46:11 -0700271 * node_dirtyable_memory - number of dirtyable pages in a node
272 * @pgdat: the node
Johannes Weinera8045522014-01-29 14:05:39 -0800273 *
Mel Gorman281e3722016-07-28 15:46:11 -0700274 * Returns the node's number of pages potentially available for dirty
275 * page cache. This is the base value for the per-node dirty limits.
Johannes Weinera8045522014-01-29 14:05:39 -0800276 */
Mel Gorman281e3722016-07-28 15:46:11 -0700277static unsigned long node_dirtyable_memory(struct pglist_data *pgdat)
Johannes Weinera8045522014-01-29 14:05:39 -0800278{
Mel Gorman281e3722016-07-28 15:46:11 -0700279 unsigned long nr_pages = 0;
280 int z;
Johannes Weinera8045522014-01-29 14:05:39 -0800281
Mel Gorman281e3722016-07-28 15:46:11 -0700282 for (z = 0; z < MAX_NR_ZONES; z++) {
283 struct zone *zone = pgdat->node_zones + z;
284
285 if (!populated_zone(zone))
286 continue;
287
288 nr_pages += zone_page_state(zone, NR_FREE_PAGES);
289 }
290
Johannes Weinera8d01432016-01-14 15:20:15 -0800291 /*
292 * Pages reserved for the kernel should not be considered
293 * dirtyable, to prevent a situation where reclaim has to
294 * clean pages in order to balance the zones.
295 */
Mel Gorman281e3722016-07-28 15:46:11 -0700296 nr_pages -= min(nr_pages, pgdat->totalreserve_pages);
Johannes Weinera8045522014-01-29 14:05:39 -0800297
Mel Gorman281e3722016-07-28 15:46:11 -0700298 nr_pages += node_page_state(pgdat, NR_INACTIVE_FILE);
299 nr_pages += node_page_state(pgdat, NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800300
301 return nr_pages;
302}
303
Johannes Weiner1edf2232012-01-10 15:06:57 -0800304static unsigned long highmem_dirtyable_memory(unsigned long total)
305{
306#ifdef CONFIG_HIGHMEM
307 int node;
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700308 unsigned long x = 0;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700309 int i;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800310
311 for_each_node_state(node, N_HIGH_MEMORY) {
Mel Gorman281e3722016-07-28 15:46:11 -0700312 for (i = ZONE_NORMAL + 1; i < MAX_NR_ZONES; i++) {
313 struct zone *z;
Minchan Kim9cb937e2016-07-28 15:47:08 -0700314 unsigned long nr_pages;
Johannes Weiner1edf2232012-01-10 15:06:57 -0800315
Mel Gorman281e3722016-07-28 15:46:11 -0700316 if (!is_highmem_idx(i))
317 continue;
318
319 z = &NODE_DATA(node)->node_zones[i];
Minchan Kim9cb937e2016-07-28 15:47:08 -0700320 if (!populated_zone(z))
321 continue;
Mel Gorman281e3722016-07-28 15:46:11 -0700322
Minchan Kim9cb937e2016-07-28 15:47:08 -0700323 nr_pages = zone_page_state(z, NR_FREE_PAGES);
Mel Gorman281e3722016-07-28 15:46:11 -0700324 /* watch for underflows */
Minchan Kim9cb937e2016-07-28 15:47:08 -0700325 nr_pages -= min(nr_pages, high_wmark_pages(z));
Mel Gormanbb4cc2b2016-07-28 15:47:29 -0700326 nr_pages += zone_page_state(z, NR_ZONE_INACTIVE_FILE);
327 nr_pages += zone_page_state(z, NR_ZONE_ACTIVE_FILE);
328 x += nr_pages;
Joonsoo Kim09b4ab32016-05-19 17:12:20 -0700329 }
Johannes Weiner1edf2232012-01-10 15:06:57 -0800330 }
Mel Gorman281e3722016-07-28 15:46:11 -0700331
Johannes Weiner1edf2232012-01-10 15:06:57 -0800332 /*
Sonny Raoc8b74c2f2012-12-20 15:05:07 -0800333 * Unreclaimable memory (kernel memory or anonymous memory
334 * without swap) can bring down the dirtyable pages below
335 * the zone's dirty balance reserve and the above calculation
336 * will underflow. However we still want to add in nodes
337 * which are below threshold (negative values) to get a more
338 * accurate calculation but make sure that the total never
339 * underflows.
340 */
341 if ((long)x < 0)
342 x = 0;
343
344 /*
Johannes Weiner1edf2232012-01-10 15:06:57 -0800345 * Make sure that the number of highmem pages is never larger
346 * than the number of the total dirtyable memory. This can only
347 * occur in very strange VM situations but we want to make sure
348 * that this does not occur.
349 */
350 return min(x, total);
351#else
352 return 0;
353#endif
354}
355
356/**
Johannes Weinerccafa282012-01-10 15:07:44 -0800357 * global_dirtyable_memory - number of globally dirtyable pages
Johannes Weiner1edf2232012-01-10 15:06:57 -0800358 *
Johannes Weinerccafa282012-01-10 15:07:44 -0800359 * Returns the global number of pages potentially available for dirty
360 * page cache. This is the base value for the global dirty limits.
Johannes Weiner1edf2232012-01-10 15:06:57 -0800361 */
H Hartley Sweeten18cf8cf2012-04-12 13:44:20 -0700362static unsigned long global_dirtyable_memory(void)
Johannes Weiner1edf2232012-01-10 15:06:57 -0800363{
364 unsigned long x;
365
Michal Hockoc41f0122017-09-06 16:23:36 -0700366 x = global_zone_page_state(NR_FREE_PAGES);
Johannes Weinera8d01432016-01-14 15:20:15 -0800367 /*
368 * Pages reserved for the kernel should not be considered
369 * dirtyable, to prevent a situation where reclaim has to
370 * clean pages in order to balance the zones.
371 */
372 x -= min(x, totalreserve_pages);
Johannes Weiner1edf2232012-01-10 15:06:57 -0800373
Mel Gorman599d0c92016-07-28 15:45:31 -0700374 x += global_node_page_state(NR_INACTIVE_FILE);
375 x += global_node_page_state(NR_ACTIVE_FILE);
Johannes Weinera8045522014-01-29 14:05:39 -0800376
Johannes Weiner1edf2232012-01-10 15:06:57 -0800377 if (!vm_highmem_is_dirtyable)
378 x -= highmem_dirtyable_memory(x);
379
380 return x + 1; /* Ensure that we never return 0 */
381}
382
Tejun Heo9fc3a432015-05-22 18:23:30 -0400383/**
384 * domain_dirty_limits - calculate thresh and bg_thresh for a wb_domain
385 * @dtc: dirty_throttle_control of interest
Johannes Weinerccafa282012-01-10 15:07:44 -0800386 *
Tejun Heo9fc3a432015-05-22 18:23:30 -0400387 * Calculate @dtc->thresh and ->bg_thresh considering
388 * vm_dirty_{bytes|ratio} and dirty_background_{bytes|ratio}. The caller
389 * must ensure that @dtc->avail is set before calling this function. The
390 * dirty limits will be lifted by 1/4 for PF_LESS_THROTTLE (ie. nfsd) and
Johannes Weinerccafa282012-01-10 15:07:44 -0800391 * real-time tasks.
392 */
Tejun Heo9fc3a432015-05-22 18:23:30 -0400393static void domain_dirty_limits(struct dirty_throttle_control *dtc)
394{
395 const unsigned long available_memory = dtc->avail;
396 struct dirty_throttle_control *gdtc = mdtc_gdtc(dtc);
397 unsigned long bytes = vm_dirty_bytes;
398 unsigned long bg_bytes = dirty_background_bytes;
Tejun Heo62a584f2016-05-27 14:34:46 -0400399 /* convert ratios to per-PAGE_SIZE for higher precision */
400 unsigned long ratio = (vm_dirty_ratio * PAGE_SIZE) / 100;
401 unsigned long bg_ratio = (dirty_background_ratio * PAGE_SIZE) / 100;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400402 unsigned long thresh;
403 unsigned long bg_thresh;
404 struct task_struct *tsk;
405
406 /* gdtc is !NULL iff @dtc is for memcg domain */
407 if (gdtc) {
408 unsigned long global_avail = gdtc->avail;
409
410 /*
411 * The byte settings can't be applied directly to memcg
412 * domains. Convert them to ratios by scaling against
Tejun Heo62a584f2016-05-27 14:34:46 -0400413 * globally available memory. As the ratios are in
414 * per-PAGE_SIZE, they can be obtained by dividing bytes by
415 * number of pages.
Tejun Heo9fc3a432015-05-22 18:23:30 -0400416 */
417 if (bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400418 ratio = min(DIV_ROUND_UP(bytes, global_avail),
419 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400420 if (bg_bytes)
Tejun Heo62a584f2016-05-27 14:34:46 -0400421 bg_ratio = min(DIV_ROUND_UP(bg_bytes, global_avail),
422 PAGE_SIZE);
Tejun Heo9fc3a432015-05-22 18:23:30 -0400423 bytes = bg_bytes = 0;
424 }
425
426 if (bytes)
427 thresh = DIV_ROUND_UP(bytes, PAGE_SIZE);
428 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400429 thresh = (ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400430
431 if (bg_bytes)
432 bg_thresh = DIV_ROUND_UP(bg_bytes, PAGE_SIZE);
433 else
Tejun Heo62a584f2016-05-27 14:34:46 -0400434 bg_thresh = (bg_ratio * available_memory) / PAGE_SIZE;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400435
Michal Hocko90daf302017-11-29 16:10:58 -0800436 if (bg_thresh >= thresh)
Tejun Heo9fc3a432015-05-22 18:23:30 -0400437 bg_thresh = thresh / 2;
438 tsk = current;
439 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk)) {
NeilBrowna53eaff2016-05-20 16:58:53 -0700440 bg_thresh += bg_thresh / 4 + global_wb_domain.dirty_limit / 32;
441 thresh += thresh / 4 + global_wb_domain.dirty_limit / 32;
Tejun Heo9fc3a432015-05-22 18:23:30 -0400442 }
443 dtc->thresh = thresh;
444 dtc->bg_thresh = bg_thresh;
445
446 /* we should eventually report the domain in the TP */
447 if (!gdtc)
448 trace_global_dirty_state(bg_thresh, thresh);
449}
450
451/**
452 * global_dirty_limits - background-writeback and dirty-throttling thresholds
453 * @pbackground: out parameter for bg_thresh
454 * @pdirty: out parameter for thresh
455 *
456 * Calculate bg_thresh and thresh for global_wb_domain. See
457 * domain_dirty_limits() for details.
458 */
Johannes Weinerccafa282012-01-10 15:07:44 -0800459void global_dirty_limits(unsigned long *pbackground, unsigned long *pdirty)
460{
Tejun Heo9fc3a432015-05-22 18:23:30 -0400461 struct dirty_throttle_control gdtc = { GDTC_INIT_NO_WB };
Johannes Weinerccafa282012-01-10 15:07:44 -0800462
Tejun Heo9fc3a432015-05-22 18:23:30 -0400463 gdtc.avail = global_dirtyable_memory();
464 domain_dirty_limits(&gdtc);
Johannes Weinerccafa282012-01-10 15:07:44 -0800465
Tejun Heo9fc3a432015-05-22 18:23:30 -0400466 *pbackground = gdtc.bg_thresh;
467 *pdirty = gdtc.thresh;
Johannes Weinerccafa282012-01-10 15:07:44 -0800468}
469
Johannes Weinera756cf52012-01-10 15:07:49 -0800470/**
Mel Gorman281e3722016-07-28 15:46:11 -0700471 * node_dirty_limit - maximum number of dirty pages allowed in a node
472 * @pgdat: the node
Johannes Weinera756cf52012-01-10 15:07:49 -0800473 *
Mel Gorman281e3722016-07-28 15:46:11 -0700474 * Returns the maximum number of dirty pages allowed in a node, based
475 * on the node's dirtyable memory.
Johannes Weinera756cf52012-01-10 15:07:49 -0800476 */
Mel Gorman281e3722016-07-28 15:46:11 -0700477static unsigned long node_dirty_limit(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800478{
Mel Gorman281e3722016-07-28 15:46:11 -0700479 unsigned long node_memory = node_dirtyable_memory(pgdat);
Johannes Weinera756cf52012-01-10 15:07:49 -0800480 struct task_struct *tsk = current;
481 unsigned long dirty;
482
483 if (vm_dirty_bytes)
484 dirty = DIV_ROUND_UP(vm_dirty_bytes, PAGE_SIZE) *
Mel Gorman281e3722016-07-28 15:46:11 -0700485 node_memory / global_dirtyable_memory();
Johannes Weinera756cf52012-01-10 15:07:49 -0800486 else
Mel Gorman281e3722016-07-28 15:46:11 -0700487 dirty = vm_dirty_ratio * node_memory / 100;
Johannes Weinera756cf52012-01-10 15:07:49 -0800488
489 if (tsk->flags & PF_LESS_THROTTLE || rt_task(tsk))
490 dirty += dirty / 4;
491
492 return dirty;
493}
494
495/**
Mel Gorman281e3722016-07-28 15:46:11 -0700496 * node_dirty_ok - tells whether a node is within its dirty limits
497 * @pgdat: the node to check
Johannes Weinera756cf52012-01-10 15:07:49 -0800498 *
Mel Gorman281e3722016-07-28 15:46:11 -0700499 * Returns %true when the dirty pages in @pgdat are within the node's
Johannes Weinera756cf52012-01-10 15:07:49 -0800500 * dirty limit, %false if the limit is exceeded.
501 */
Mel Gorman281e3722016-07-28 15:46:11 -0700502bool node_dirty_ok(struct pglist_data *pgdat)
Johannes Weinera756cf52012-01-10 15:07:49 -0800503{
Mel Gorman281e3722016-07-28 15:46:11 -0700504 unsigned long limit = node_dirty_limit(pgdat);
505 unsigned long nr_pages = 0;
Johannes Weinera756cf52012-01-10 15:07:49 -0800506
Mel Gorman11fb9982016-07-28 15:46:20 -0700507 nr_pages += node_page_state(pgdat, NR_FILE_DIRTY);
508 nr_pages += node_page_state(pgdat, NR_UNSTABLE_NFS);
509 nr_pages += node_page_state(pgdat, NR_WRITEBACK);
Mel Gorman281e3722016-07-28 15:46:11 -0700510
511 return nr_pages <= limit;
Johannes Weinera756cf52012-01-10 15:07:49 -0800512}
513
David Rientjes2da02992009-01-06 14:39:31 -0800514int dirty_background_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700515 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800516 loff_t *ppos)
517{
518 int ret;
519
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700520 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800521 if (ret == 0 && write)
522 dirty_background_bytes = 0;
523 return ret;
524}
525
526int dirty_background_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700527 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800528 loff_t *ppos)
529{
530 int ret;
531
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700532 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800533 if (ret == 0 && write)
534 dirty_background_ratio = 0;
535 return ret;
536}
537
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700538int dirty_ratio_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700539 void __user *buffer, size_t *lenp,
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700540 loff_t *ppos)
541{
542 int old_ratio = vm_dirty_ratio;
David Rientjes2da02992009-01-06 14:39:31 -0800543 int ret;
544
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700545 ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700546 if (ret == 0 && write && vm_dirty_ratio != old_ratio) {
Jan Karaeb608e32012-05-24 18:59:11 +0200547 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800548 vm_dirty_bytes = 0;
549 }
550 return ret;
551}
552
David Rientjes2da02992009-01-06 14:39:31 -0800553int dirty_bytes_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700554 void __user *buffer, size_t *lenp,
David Rientjes2da02992009-01-06 14:39:31 -0800555 loff_t *ppos)
556{
Sven Wegenerfc3501d2009-02-11 13:04:23 -0800557 unsigned long old_bytes = vm_dirty_bytes;
David Rientjes2da02992009-01-06 14:39:31 -0800558 int ret;
559
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700560 ret = proc_doulongvec_minmax(table, write, buffer, lenp, ppos);
David Rientjes2da02992009-01-06 14:39:31 -0800561 if (ret == 0 && write && vm_dirty_bytes != old_bytes) {
Jan Karaeb608e32012-05-24 18:59:11 +0200562 writeback_set_ratelimit();
David Rientjes2da02992009-01-06 14:39:31 -0800563 vm_dirty_ratio = 0;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700564 }
565 return ret;
566}
567
Jan Karaeb608e32012-05-24 18:59:11 +0200568static unsigned long wp_next_time(unsigned long cur_time)
569{
570 cur_time += VM_COMPLETIONS_PERIOD_LEN;
571 /* 0 has a special meaning... */
572 if (!cur_time)
573 return 1;
574 return cur_time;
575}
576
Tejun Heoc7981432015-05-22 18:23:29 -0400577static void wb_domain_writeout_inc(struct wb_domain *dom,
578 struct fprop_local_percpu *completions,
579 unsigned int max_prop_frac)
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700580{
Tejun Heoc7981432015-05-22 18:23:29 -0400581 __fprop_inc_percpu_max(&dom->completions, completions,
582 max_prop_frac);
Jan Karaeb608e32012-05-24 18:59:11 +0200583 /* First event after period switching was turned off? */
Steven Rostedt (VMware)517663e2017-02-24 14:59:24 -0800584 if (unlikely(!dom->period_time)) {
Jan Karaeb608e32012-05-24 18:59:11 +0200585 /*
586 * We can race with other __bdi_writeout_inc calls here but
587 * it does not cause any harm since the resulting time when
588 * timer will fire and what is in writeout_period_time will be
589 * roughly the same.
590 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400591 dom->period_time = wp_next_time(jiffies);
592 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200593 }
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700594}
595
Tejun Heoc7981432015-05-22 18:23:29 -0400596/*
597 * Increment @wb's writeout completion count and the global writeout
598 * completion count. Called from test_clear_page_writeback().
599 */
600static inline void __wb_writeout_inc(struct bdi_writeback *wb)
601{
Tejun Heo841710a2015-05-22 18:23:33 -0400602 struct wb_domain *cgdom;
603
Nikolay Borisov3e8f3992017-07-12 14:37:51 -0700604 inc_wb_stat(wb, WB_WRITTEN);
Tejun Heoc7981432015-05-22 18:23:29 -0400605 wb_domain_writeout_inc(&global_wb_domain, &wb->completions,
606 wb->bdi->max_prop_frac);
Tejun Heo841710a2015-05-22 18:23:33 -0400607
608 cgdom = mem_cgroup_wb_domain(wb);
609 if (cgdom)
610 wb_domain_writeout_inc(cgdom, wb_memcg_completions(wb),
611 wb->bdi->max_prop_frac);
Tejun Heoc7981432015-05-22 18:23:29 -0400612}
613
Tejun Heo93f78d82015-05-22 17:13:27 -0400614void wb_writeout_inc(struct bdi_writeback *wb)
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700615{
616 unsigned long flags;
617
618 local_irq_save(flags);
Tejun Heo93f78d82015-05-22 17:13:27 -0400619 __wb_writeout_inc(wb);
Miklos Szeredidd5656e2008-04-30 00:54:37 -0700620 local_irq_restore(flags);
621}
Tejun Heo93f78d82015-05-22 17:13:27 -0400622EXPORT_SYMBOL_GPL(wb_writeout_inc);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700623
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700624/*
Jan Karaeb608e32012-05-24 18:59:11 +0200625 * On idle system, we can be called long after we scheduled because we use
626 * deferred timers so count with missed periods.
627 */
Kees Cook9823e512017-11-15 17:37:04 -0800628static void writeout_period(struct timer_list *t)
Jan Karaeb608e32012-05-24 18:59:11 +0200629{
Kees Cook9823e512017-11-15 17:37:04 -0800630 struct wb_domain *dom = from_timer(dom, t, period_timer);
Tejun Heo380c27c2015-05-22 18:23:21 -0400631 int miss_periods = (jiffies - dom->period_time) /
Jan Karaeb608e32012-05-24 18:59:11 +0200632 VM_COMPLETIONS_PERIOD_LEN;
633
Tejun Heo380c27c2015-05-22 18:23:21 -0400634 if (fprop_new_period(&dom->completions, miss_periods + 1)) {
635 dom->period_time = wp_next_time(dom->period_time +
Jan Karaeb608e32012-05-24 18:59:11 +0200636 miss_periods * VM_COMPLETIONS_PERIOD_LEN);
Tejun Heo380c27c2015-05-22 18:23:21 -0400637 mod_timer(&dom->period_timer, dom->period_time);
Jan Karaeb608e32012-05-24 18:59:11 +0200638 } else {
639 /*
640 * Aging has zeroed all fractions. Stop wasting CPU on period
641 * updates.
642 */
Tejun Heo380c27c2015-05-22 18:23:21 -0400643 dom->period_time = 0;
Jan Karaeb608e32012-05-24 18:59:11 +0200644 }
645}
646
Tejun Heo380c27c2015-05-22 18:23:21 -0400647int wb_domain_init(struct wb_domain *dom, gfp_t gfp)
648{
649 memset(dom, 0, sizeof(*dom));
Tejun Heodcc25ae2015-05-22 18:23:22 -0400650
651 spin_lock_init(&dom->lock);
652
Kees Cook9823e512017-11-15 17:37:04 -0800653 timer_setup(&dom->period_timer, writeout_period, TIMER_DEFERRABLE);
Tejun Heodcc25ae2015-05-22 18:23:22 -0400654
655 dom->dirty_limit_tstamp = jiffies;
656
Tejun Heo380c27c2015-05-22 18:23:21 -0400657 return fprop_global_init(&dom->completions, gfp);
658}
659
Tejun Heo841710a2015-05-22 18:23:33 -0400660#ifdef CONFIG_CGROUP_WRITEBACK
661void wb_domain_exit(struct wb_domain *dom)
662{
663 del_timer_sync(&dom->period_timer);
664 fprop_global_destroy(&dom->completions);
665}
666#endif
667
Jan Karaeb608e32012-05-24 18:59:11 +0200668/*
Johannes Weinerd08c4292011-10-31 17:07:05 -0700669 * bdi_min_ratio keeps the sum of the minimum dirty shares of all
670 * registered backing devices, which, for obvious reasons, can not
671 * exceed 100%.
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700672 */
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700673static unsigned int bdi_min_ratio;
674
675int bdi_set_min_ratio(struct backing_dev_info *bdi, unsigned int min_ratio)
676{
677 int ret = 0;
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700678
Jens Axboecfc4ba52009-09-14 13:12:40 +0200679 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700680 if (min_ratio > bdi->max_ratio) {
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700681 ret = -EINVAL;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700682 } else {
683 min_ratio -= bdi->min_ratio;
684 if (bdi_min_ratio + min_ratio < 100) {
685 bdi_min_ratio += min_ratio;
686 bdi->min_ratio += min_ratio;
687 } else {
688 ret = -EINVAL;
689 }
690 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200691 spin_unlock_bh(&bdi_lock);
Peter Zijlstra189d3c42008-04-30 00:54:35 -0700692
693 return ret;
694}
695
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700696int bdi_set_max_ratio(struct backing_dev_info *bdi, unsigned max_ratio)
697{
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700698 int ret = 0;
699
700 if (max_ratio > 100)
701 return -EINVAL;
702
Jens Axboecfc4ba52009-09-14 13:12:40 +0200703 spin_lock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700704 if (bdi->min_ratio > max_ratio) {
705 ret = -EINVAL;
706 } else {
707 bdi->max_ratio = max_ratio;
Jan Karaeb608e32012-05-24 18:59:11 +0200708 bdi->max_prop_frac = (FPROP_FRAC_BASE * max_ratio) / 100;
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700709 }
Jens Axboecfc4ba52009-09-14 13:12:40 +0200710 spin_unlock_bh(&bdi_lock);
Peter Zijlstraa42dde02008-04-30 00:54:36 -0700711
712 return ret;
713}
714EXPORT_SYMBOL(bdi_set_max_ratio);
715
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600716static unsigned long dirty_freerun_ceiling(unsigned long thresh,
717 unsigned long bg_thresh)
718{
719 return (thresh + bg_thresh) / 2;
720}
721
Tejun Heoc7981432015-05-22 18:23:29 -0400722static unsigned long hard_dirty_limit(struct wb_domain *dom,
723 unsigned long thresh)
Wu Fengguangffd1f602011-06-19 22:18:42 -0600724{
Tejun Heodcc25ae2015-05-22 18:23:22 -0400725 return max(thresh, dom->dirty_limit);
Wu Fengguangffd1f602011-06-19 22:18:42 -0600726}
727
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400728/*
729 * Memory which can be further allocated to a memcg domain is capped by
730 * system-wide clean memory excluding the amount being used in the domain.
731 */
732static void mdtc_calc_avail(struct dirty_throttle_control *mdtc,
733 unsigned long filepages, unsigned long headroom)
Tejun Heoc2aa7232015-05-22 18:23:35 -0400734{
735 struct dirty_throttle_control *gdtc = mdtc_gdtc(mdtc);
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400736 unsigned long clean = filepages - min(filepages, mdtc->dirty);
737 unsigned long global_clean = gdtc->avail - min(gdtc->avail, gdtc->dirty);
738 unsigned long other_clean = global_clean - min(global_clean, clean);
Tejun Heoc2aa7232015-05-22 18:23:35 -0400739
Tejun Heoc5edf9c2015-09-29 13:04:26 -0400740 mdtc->avail = filepages + min(headroom, other_clean);
Christoph Lameter1b424462007-05-06 14:48:59 -0700741}
742
Wu Fengguang6f718652011-03-02 17:14:34 -0600743/**
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400744 * __wb_calc_thresh - @wb's share of dirty throttling threshold
745 * @dtc: dirty_throttle_context of interest
Wu Fengguang1babe182010-08-11 14:17:40 -0700746 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400747 * Returns @wb's dirty limit in pages. The term "dirty" in the context of
Wu Fengguang6f718652011-03-02 17:14:34 -0600748 * dirty balancing includes all PG_dirty, PG_writeback and NFS unstable pages.
Wu Fengguangaed21ad2011-11-23 11:44:41 -0600749 *
750 * Note that balance_dirty_pages() will only seriously take it as a hard limit
751 * when sleeping max_pause per page is not enough to keep the dirty pages under
752 * control. For example, when the device is completely stalled due to some error
753 * conditions, or when there are 1000 dd tasks writing to a slow 10MB/s USB key.
754 * In the other normal situations, it acts more gently by throttling the tasks
Tejun Heoa88a3412015-05-22 17:13:28 -0400755 * more (rather than completely block them) when the wb dirty pages go high.
Wu Fengguang6f718652011-03-02 17:14:34 -0600756 *
757 * It allocates high/low dirty limits to fast/slow devices, in order to prevent
Wu Fengguang1babe182010-08-11 14:17:40 -0700758 * - starving fast devices
759 * - piling up dirty pages (that will take long time to sync) on slow devices
760 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400761 * The wb's share of dirty limit will be adapting to its throughput and
Wu Fengguang1babe182010-08-11 14:17:40 -0700762 * bounded by the bdi->min_ratio and/or bdi->max_ratio parameters, if set.
763 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400764static unsigned long __wb_calc_thresh(struct dirty_throttle_control *dtc)
Wu Fengguang16c40422010-08-11 14:17:39 -0700765{
Tejun Heoe9f07df2015-05-22 18:23:28 -0400766 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400767 unsigned long thresh = dtc->thresh;
Tejun Heo0d960a32015-05-22 18:23:19 -0400768 u64 wb_thresh;
Wu Fengguang16c40422010-08-11 14:17:39 -0700769 long numerator, denominator;
Tejun Heo693108a2015-05-22 17:13:49 -0400770 unsigned long wb_min_ratio, wb_max_ratio;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700771
Wu Fengguang16c40422010-08-11 14:17:39 -0700772 /*
Tejun Heo0d960a32015-05-22 18:23:19 -0400773 * Calculate this BDI's share of the thresh ratio.
Wu Fengguang16c40422010-08-11 14:17:39 -0700774 */
Tejun Heoe9770b32015-05-22 18:23:27 -0400775 fprop_fraction_percpu(&dom->completions, dtc->wb_completions,
Tejun Heo380c27c2015-05-22 18:23:21 -0400776 &numerator, &denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700777
Tejun Heo0d960a32015-05-22 18:23:19 -0400778 wb_thresh = (thresh * (100 - bdi_min_ratio)) / 100;
779 wb_thresh *= numerator;
780 do_div(wb_thresh, denominator);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -0700781
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400782 wb_min_max_ratio(dtc->wb, &wb_min_ratio, &wb_max_ratio);
Wu Fengguang16c40422010-08-11 14:17:39 -0700783
Tejun Heo0d960a32015-05-22 18:23:19 -0400784 wb_thresh += (thresh * wb_min_ratio) / 100;
785 if (wb_thresh > (thresh * wb_max_ratio) / 100)
786 wb_thresh = thresh * wb_max_ratio / 100;
Wu Fengguang16c40422010-08-11 14:17:39 -0700787
Tejun Heo0d960a32015-05-22 18:23:19 -0400788 return wb_thresh;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789}
790
Tejun Heob1cbc6d2015-05-22 18:23:25 -0400791unsigned long wb_calc_thresh(struct bdi_writeback *wb, unsigned long thresh)
792{
793 struct dirty_throttle_control gdtc = { GDTC_INIT(wb),
794 .thresh = thresh };
795 return __wb_calc_thresh(&gdtc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796}
797
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600798/*
Maxim Patlasov5a537482013-09-11 14:22:46 -0700799 * setpoint - dirty 3
800 * f(dirty) := 1.0 + (----------------)
801 * limit - setpoint
802 *
803 * it's a 3rd order polynomial that subjects to
804 *
805 * (1) f(freerun) = 2.0 => rampup dirty_ratelimit reasonably fast
806 * (2) f(setpoint) = 1.0 => the balance point
807 * (3) f(limit) = 0 => the hard limit
808 * (4) df/dx <= 0 => negative feedback control
809 * (5) the closer to setpoint, the smaller |df/dx| (and the reverse)
810 * => fast response on large errors; small oscillation near setpoint
811 */
Rik van Rield5c9fde2014-05-06 12:50:01 -0700812static long long pos_ratio_polynom(unsigned long setpoint,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700813 unsigned long dirty,
814 unsigned long limit)
815{
816 long long pos_ratio;
817 long x;
818
Rik van Rield5c9fde2014-05-06 12:50:01 -0700819 x = div64_s64(((s64)setpoint - (s64)dirty) << RATELIMIT_CALC_SHIFT,
Tejun Heo464d1382015-04-21 16:49:13 -0400820 (limit - setpoint) | 1);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700821 pos_ratio = x;
822 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
823 pos_ratio = pos_ratio * x >> RATELIMIT_CALC_SHIFT;
824 pos_ratio += 1 << RATELIMIT_CALC_SHIFT;
825
826 return clamp(pos_ratio, 0LL, 2LL << RATELIMIT_CALC_SHIFT);
827}
828
829/*
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600830 * Dirty position control.
831 *
832 * (o) global/bdi setpoints
833 *
Tejun Heode1fff32015-05-22 17:13:29 -0400834 * We want the dirty pages be balanced around the global/wb setpoints.
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600835 * When the number of dirty pages is higher/lower than the setpoint, the
836 * dirty position control ratio (and hence task dirty ratelimit) will be
837 * decreased/increased to bring the dirty pages back to the setpoint.
838 *
839 * pos_ratio = 1 << RATELIMIT_CALC_SHIFT
840 *
841 * if (dirty < setpoint) scale up pos_ratio
842 * if (dirty > setpoint) scale down pos_ratio
843 *
Tejun Heode1fff32015-05-22 17:13:29 -0400844 * if (wb_dirty < wb_setpoint) scale up pos_ratio
845 * if (wb_dirty > wb_setpoint) scale down pos_ratio
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600846 *
847 * task_ratelimit = dirty_ratelimit * pos_ratio >> RATELIMIT_CALC_SHIFT
848 *
849 * (o) global control line
850 *
851 * ^ pos_ratio
852 * |
853 * | |<===== global dirty control scope ======>|
854 * 2.0 .............*
855 * | .*
856 * | . *
857 * | . *
858 * | . *
859 * | . *
860 * | . *
861 * 1.0 ................................*
862 * | . . *
863 * | . . *
864 * | . . *
865 * | . . *
866 * | . . *
867 * 0 +------------.------------------.----------------------*------------->
868 * freerun^ setpoint^ limit^ dirty pages
869 *
Tejun Heode1fff32015-05-22 17:13:29 -0400870 * (o) wb control line
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600871 *
872 * ^ pos_ratio
873 * |
874 * | *
875 * | *
876 * | *
877 * | *
878 * | * |<=========== span ============>|
879 * 1.0 .......................*
880 * | . *
881 * | . *
882 * | . *
883 * | . *
884 * | . *
885 * | . *
886 * | . *
887 * | . *
888 * | . *
889 * | . *
890 * | . *
891 * 1/4 ...............................................* * * * * * * * * * * *
892 * | . .
893 * | . .
894 * | . .
895 * 0 +----------------------.-------------------------------.------------->
Tejun Heode1fff32015-05-22 17:13:29 -0400896 * wb_setpoint^ x_intercept^
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600897 *
Tejun Heode1fff32015-05-22 17:13:29 -0400898 * The wb control line won't drop below pos_ratio=1/4, so that wb_dirty can
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600899 * be smoothly throttled down to normal if it starts high in situations like
900 * - start writing to a slow SD card and a fast disk at the same time. The SD
Tejun Heode1fff32015-05-22 17:13:29 -0400901 * card's wb_dirty may rush to many times higher than wb_setpoint.
902 * - the wb dirty thresh drops quickly due to change of JBOD workload
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600903 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400904static void wb_position_ratio(struct dirty_throttle_control *dtc)
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600905{
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400906 struct bdi_writeback *wb = dtc->wb;
Tejun Heoa88a3412015-05-22 17:13:28 -0400907 unsigned long write_bw = wb->avg_write_bandwidth;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400908 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -0400909 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400910 unsigned long wb_thresh = dtc->wb_thresh;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600911 unsigned long x_intercept;
912 unsigned long setpoint; /* dirty pages' target balance point */
Tejun Heode1fff32015-05-22 17:13:29 -0400913 unsigned long wb_setpoint;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600914 unsigned long span;
915 long long pos_ratio; /* for scaling up/down the rate limit */
916 long x;
917
Tejun Heodaddfa32015-05-22 18:23:26 -0400918 dtc->pos_ratio = 0;
919
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400920 if (unlikely(dtc->dirty >= limit))
Tejun Heodaddfa32015-05-22 18:23:26 -0400921 return;
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600922
923 /*
924 * global setpoint
925 *
Maxim Patlasov5a537482013-09-11 14:22:46 -0700926 * See comment for pos_ratio_polynom().
Wu Fengguang6c14ae12011-03-02 16:04:18 -0600927 */
928 setpoint = (freerun + limit) / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400929 pos_ratio = pos_ratio_polynom(setpoint, dtc->dirty, limit);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700930
931 /*
932 * The strictlimit feature is a tool preventing mistrusted filesystems
933 * from growing a large number of dirty pages before throttling. For
Tejun Heode1fff32015-05-22 17:13:29 -0400934 * such filesystems balance_dirty_pages always checks wb counters
935 * against wb limits. Even if global "nr_dirty" is under "freerun".
Maxim Patlasov5a537482013-09-11 14:22:46 -0700936 * This is especially important for fuse which sets bdi->max_ratio to
937 * 1% by default. Without strictlimit feature, fuse writeback may
938 * consume arbitrary amount of RAM because it is accounted in
939 * NR_WRITEBACK_TEMP which is not involved in calculating "nr_dirty".
940 *
Tejun Heoa88a3412015-05-22 17:13:28 -0400941 * Here, in wb_position_ratio(), we calculate pos_ratio based on
Tejun Heode1fff32015-05-22 17:13:29 -0400942 * two values: wb_dirty and wb_thresh. Let's consider an example:
Maxim Patlasov5a537482013-09-11 14:22:46 -0700943 * total amount of RAM is 16GB, bdi->max_ratio is equal to 1%, global
944 * limits are set by default to 10% and 20% (background and throttle).
Tejun Heode1fff32015-05-22 17:13:29 -0400945 * Then wb_thresh is 1% of 20% of 16GB. This amounts to ~8K pages.
Tejun Heo0d960a32015-05-22 18:23:19 -0400946 * wb_calc_thresh(wb, bg_thresh) is about ~4K pages. wb_setpoint is
Tejun Heode1fff32015-05-22 17:13:29 -0400947 * about ~6K pages (as the average of background and throttle wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700948 * limits). The 3rd order polynomial will provide positive feedback if
Tejun Heode1fff32015-05-22 17:13:29 -0400949 * wb_dirty is under wb_setpoint and vice versa.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700950 *
951 * Note, that we cannot use global counters in these calculations
Tejun Heode1fff32015-05-22 17:13:29 -0400952 * because we want to throttle process writing to a strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700953 * much earlier than global "freerun" is reached (~23MB vs. ~2.3GB
954 * in the example above).
955 */
Tejun Heoa88a3412015-05-22 17:13:28 -0400956 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heode1fff32015-05-22 17:13:29 -0400957 long long wb_pos_ratio;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700958
Tejun Heodaddfa32015-05-22 18:23:26 -0400959 if (dtc->wb_dirty < 8) {
960 dtc->pos_ratio = min_t(long long, pos_ratio * 2,
961 2 << RATELIMIT_CALC_SHIFT);
962 return;
963 }
Maxim Patlasov5a537482013-09-11 14:22:46 -0700964
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400965 if (dtc->wb_dirty >= wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400966 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700967
Tejun Heo970fb012015-05-22 18:23:24 -0400968 wb_setpoint = dirty_freerun_ceiling(wb_thresh,
969 dtc->wb_bg_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700970
Tejun Heode1fff32015-05-22 17:13:29 -0400971 if (wb_setpoint == 0 || wb_setpoint == wb_thresh)
Tejun Heodaddfa32015-05-22 18:23:26 -0400972 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -0700973
Tejun Heo2bc00ae2015-05-22 18:23:23 -0400974 wb_pos_ratio = pos_ratio_polynom(wb_setpoint, dtc->wb_dirty,
Tejun Heode1fff32015-05-22 17:13:29 -0400975 wb_thresh);
Maxim Patlasov5a537482013-09-11 14:22:46 -0700976
977 /*
Tejun Heode1fff32015-05-22 17:13:29 -0400978 * Typically, for strictlimit case, wb_setpoint << setpoint
979 * and pos_ratio >> wb_pos_ratio. In the other words global
Maxim Patlasov5a537482013-09-11 14:22:46 -0700980 * state ("dirty") is not limiting factor and we have to
Tejun Heode1fff32015-05-22 17:13:29 -0400981 * make decision based on wb counters. But there is an
Maxim Patlasov5a537482013-09-11 14:22:46 -0700982 * important case when global pos_ratio should get precedence:
983 * global limits are exceeded (e.g. due to activities on other
Tejun Heode1fff32015-05-22 17:13:29 -0400984 * wb's) while given strictlimit wb is below limit.
Maxim Patlasov5a537482013-09-11 14:22:46 -0700985 *
Tejun Heode1fff32015-05-22 17:13:29 -0400986 * "pos_ratio * wb_pos_ratio" would work for the case above,
Maxim Patlasov5a537482013-09-11 14:22:46 -0700987 * but it would look too non-natural for the case of all
Tejun Heode1fff32015-05-22 17:13:29 -0400988 * activity in the system coming from a single strictlimit wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700989 * with bdi->max_ratio == 100%.
990 *
991 * Note that min() below somewhat changes the dynamics of the
992 * control system. Normally, pos_ratio value can be well over 3
Tejun Heode1fff32015-05-22 17:13:29 -0400993 * (when globally we are at freerun and wb is well below wb
Maxim Patlasov5a537482013-09-11 14:22:46 -0700994 * setpoint). Now the maximum pos_ratio in the same situation
995 * is 2. We might want to tweak this if we observe the control
996 * system is too slow to adapt.
997 */
Tejun Heodaddfa32015-05-22 18:23:26 -0400998 dtc->pos_ratio = min(pos_ratio, wb_pos_ratio);
999 return;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001000 }
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001001
1002 /*
1003 * We have computed basic pos_ratio above based on global situation. If
Tejun Heode1fff32015-05-22 17:13:29 -04001004 * the wb is over/under its share of dirty pages, we want to scale
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001005 * pos_ratio further down/up. That is done by the following mechanism.
1006 */
1007
1008 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001009 * wb setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001010 *
Tejun Heode1fff32015-05-22 17:13:29 -04001011 * f(wb_dirty) := 1.0 + k * (wb_dirty - wb_setpoint)
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001012 *
Tejun Heode1fff32015-05-22 17:13:29 -04001013 * x_intercept - wb_dirty
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001014 * := --------------------------
Tejun Heode1fff32015-05-22 17:13:29 -04001015 * x_intercept - wb_setpoint
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001016 *
Tejun Heode1fff32015-05-22 17:13:29 -04001017 * The main wb control line is a linear function that subjects to
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001018 *
Tejun Heode1fff32015-05-22 17:13:29 -04001019 * (1) f(wb_setpoint) = 1.0
1020 * (2) k = - 1 / (8 * write_bw) (in single wb case)
1021 * or equally: x_intercept = wb_setpoint + 8 * write_bw
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001022 *
Tejun Heode1fff32015-05-22 17:13:29 -04001023 * For single wb case, the dirty pages are observed to fluctuate
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001024 * regularly within range
Tejun Heode1fff32015-05-22 17:13:29 -04001025 * [wb_setpoint - write_bw/2, wb_setpoint + write_bw/2]
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001026 * for various filesystems, where (2) can yield in a reasonable 12.5%
1027 * fluctuation range for pos_ratio.
1028 *
Tejun Heode1fff32015-05-22 17:13:29 -04001029 * For JBOD case, wb_thresh (not wb_dirty!) could fluctuate up to its
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001030 * own size, so move the slope over accordingly and choose a slope that
Tejun Heode1fff32015-05-22 17:13:29 -04001031 * yields 100% pos_ratio fluctuation on suddenly doubled wb_thresh.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001032 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001033 if (unlikely(wb_thresh > dtc->thresh))
1034 wb_thresh = dtc->thresh;
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001035 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001036 * It's very possible that wb_thresh is close to 0 not because the
Wu Fengguangaed21ad2011-11-23 11:44:41 -06001037 * device is slow, but that it has remained inactive for long time.
1038 * Honour such devices a reasonable good (hopefully IO efficient)
1039 * threshold, so that the occasional writes won't be blocked and active
1040 * writes can rampup the threshold quickly.
1041 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001042 wb_thresh = max(wb_thresh, (limit - dtc->dirty) / 8);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001043 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001044 * scale global setpoint to wb's:
1045 * wb_setpoint = setpoint * wb_thresh / thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001046 */
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001047 x = div_u64((u64)wb_thresh << 16, dtc->thresh | 1);
Tejun Heode1fff32015-05-22 17:13:29 -04001048 wb_setpoint = setpoint * (u64)x >> 16;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001049 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001050 * Use span=(8*write_bw) in single wb case as indicated by
1051 * (thresh - wb_thresh ~= 0) and transit to wb_thresh in JBOD case.
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001052 *
Tejun Heode1fff32015-05-22 17:13:29 -04001053 * wb_thresh thresh - wb_thresh
1054 * span = --------- * (8 * write_bw) + ------------------ * wb_thresh
1055 * thresh thresh
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001056 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001057 span = (dtc->thresh - wb_thresh + 8 * write_bw) * (u64)x >> 16;
Tejun Heode1fff32015-05-22 17:13:29 -04001058 x_intercept = wb_setpoint + span;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001059
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001060 if (dtc->wb_dirty < x_intercept - span / 4) {
1061 pos_ratio = div64_u64(pos_ratio * (x_intercept - dtc->wb_dirty),
Linus Torvaldse4bc13a2015-06-25 16:00:17 -07001062 (x_intercept - wb_setpoint) | 1);
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001063 } else
1064 pos_ratio /= 4;
1065
Wu Fengguang8927f662011-08-04 22:16:46 -06001066 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001067 * wb reserve area, safeguard against dirty pool underrun and disk idle
Wu Fengguang8927f662011-08-04 22:16:46 -06001068 * It may push the desired control point of global dirty pages higher
1069 * than setpoint.
1070 */
Tejun Heode1fff32015-05-22 17:13:29 -04001071 x_intercept = wb_thresh / 2;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001072 if (dtc->wb_dirty < x_intercept) {
1073 if (dtc->wb_dirty > x_intercept / 8)
1074 pos_ratio = div_u64(pos_ratio * x_intercept,
1075 dtc->wb_dirty);
Wu Fengguang50657fc2011-10-11 17:06:33 -06001076 else
Wu Fengguang8927f662011-08-04 22:16:46 -06001077 pos_ratio *= 8;
1078 }
1079
Tejun Heodaddfa32015-05-22 18:23:26 -04001080 dtc->pos_ratio = pos_ratio;
Wu Fengguang6c14ae12011-03-02 16:04:18 -06001081}
1082
Tejun Heoa88a3412015-05-22 17:13:28 -04001083static void wb_update_write_bandwidth(struct bdi_writeback *wb,
1084 unsigned long elapsed,
1085 unsigned long written)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001086{
1087 const unsigned long period = roundup_pow_of_two(3 * HZ);
Tejun Heoa88a3412015-05-22 17:13:28 -04001088 unsigned long avg = wb->avg_write_bandwidth;
1089 unsigned long old = wb->write_bandwidth;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001090 u64 bw;
1091
1092 /*
1093 * bw = written * HZ / elapsed
1094 *
1095 * bw * elapsed + write_bandwidth * (period - elapsed)
1096 * write_bandwidth = ---------------------------------------------------
1097 * period
Tejun Heoc72efb62015-03-23 00:18:48 -04001098 *
1099 * @written may have decreased due to account_page_redirty().
1100 * Avoid underflowing @bw calculation.
Wu Fengguange98be2d2010-08-29 11:22:30 -06001101 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001102 bw = written - min(written, wb->written_stamp);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001103 bw *= HZ;
1104 if (unlikely(elapsed > period)) {
1105 do_div(bw, elapsed);
1106 avg = bw;
1107 goto out;
1108 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001109 bw += (u64)wb->write_bandwidth * (period - elapsed);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001110 bw >>= ilog2(period);
1111
1112 /*
1113 * one more level of smoothing, for filtering out sudden spikes
1114 */
1115 if (avg > old && old >= (unsigned long)bw)
1116 avg -= (avg - old) >> 3;
1117
1118 if (avg < old && old <= (unsigned long)bw)
1119 avg += (old - avg) >> 3;
1120
1121out:
Tejun Heo95a46c62015-05-22 17:13:47 -04001122 /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
1123 avg = max(avg, 1LU);
1124 if (wb_has_dirty_io(wb)) {
1125 long delta = avg - wb->avg_write_bandwidth;
1126 WARN_ON_ONCE(atomic_long_add_return(delta,
1127 &wb->bdi->tot_write_bandwidth) <= 0);
1128 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001129 wb->write_bandwidth = bw;
1130 wb->avg_write_bandwidth = avg;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001131}
1132
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001133static void update_dirty_limit(struct dirty_throttle_control *dtc)
Wu Fengguangc42843f2011-03-02 15:54:09 -06001134{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001135 struct wb_domain *dom = dtc_dom(dtc);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001136 unsigned long thresh = dtc->thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04001137 unsigned long limit = dom->dirty_limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001138
1139 /*
1140 * Follow up in one step.
1141 */
1142 if (limit < thresh) {
1143 limit = thresh;
1144 goto update;
1145 }
1146
1147 /*
1148 * Follow down slowly. Use the higher one as the target, because thresh
1149 * may drop below dirty. This is exactly the reason to introduce
Tejun Heodcc25ae2015-05-22 18:23:22 -04001150 * dom->dirty_limit which is guaranteed to lie above the dirty pages.
Wu Fengguangc42843f2011-03-02 15:54:09 -06001151 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001152 thresh = max(thresh, dtc->dirty);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001153 if (limit > thresh) {
1154 limit -= (limit - thresh) >> 5;
1155 goto update;
1156 }
1157 return;
1158update:
Tejun Heodcc25ae2015-05-22 18:23:22 -04001159 dom->dirty_limit = limit;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001160}
1161
Tejun Heoe9f07df2015-05-22 18:23:28 -04001162static void domain_update_bandwidth(struct dirty_throttle_control *dtc,
Wu Fengguangc42843f2011-03-02 15:54:09 -06001163 unsigned long now)
1164{
Tejun Heoe9f07df2015-05-22 18:23:28 -04001165 struct wb_domain *dom = dtc_dom(dtc);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001166
1167 /*
1168 * check locklessly first to optimize away locking for the most time
1169 */
Tejun Heodcc25ae2015-05-22 18:23:22 -04001170 if (time_before(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL))
Wu Fengguangc42843f2011-03-02 15:54:09 -06001171 return;
1172
Tejun Heodcc25ae2015-05-22 18:23:22 -04001173 spin_lock(&dom->lock);
1174 if (time_after_eq(now, dom->dirty_limit_tstamp + BANDWIDTH_INTERVAL)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001175 update_dirty_limit(dtc);
Tejun Heodcc25ae2015-05-22 18:23:22 -04001176 dom->dirty_limit_tstamp = now;
Wu Fengguangc42843f2011-03-02 15:54:09 -06001177 }
Tejun Heodcc25ae2015-05-22 18:23:22 -04001178 spin_unlock(&dom->lock);
Wu Fengguangc42843f2011-03-02 15:54:09 -06001179}
1180
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001181/*
Tejun Heode1fff32015-05-22 17:13:29 -04001182 * Maintain wb->dirty_ratelimit, the base dirty throttle rate.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001183 *
Tejun Heode1fff32015-05-22 17:13:29 -04001184 * Normal wb tasks will be curbed at or below it in long term.
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001185 * Obviously it should be around (write_bw / N) when there are N dd tasks.
1186 */
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001187static void wb_update_dirty_ratelimit(struct dirty_throttle_control *dtc,
Tejun Heoa88a3412015-05-22 17:13:28 -04001188 unsigned long dirtied,
1189 unsigned long elapsed)
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001190{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001191 struct bdi_writeback *wb = dtc->wb;
1192 unsigned long dirty = dtc->dirty;
1193 unsigned long freerun = dirty_freerun_ceiling(dtc->thresh, dtc->bg_thresh);
Tejun Heoc7981432015-05-22 18:23:29 -04001194 unsigned long limit = hard_dirty_limit(dtc_dom(dtc), dtc->thresh);
Wu Fengguang73811312011-08-26 15:53:24 -06001195 unsigned long setpoint = (freerun + limit) / 2;
Tejun Heoa88a3412015-05-22 17:13:28 -04001196 unsigned long write_bw = wb->avg_write_bandwidth;
1197 unsigned long dirty_ratelimit = wb->dirty_ratelimit;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001198 unsigned long dirty_rate;
1199 unsigned long task_ratelimit;
1200 unsigned long balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001201 unsigned long step;
1202 unsigned long x;
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001203 unsigned long shift;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001204
1205 /*
1206 * The dirty rate will match the writeout rate in long term, except
1207 * when dirty pages are truncated by userspace or re-dirtied by FS.
1208 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001209 dirty_rate = (dirtied - wb->dirtied_stamp) * HZ / elapsed;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001210
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001211 /*
1212 * task_ratelimit reflects each dd's dirty rate for the past 200ms.
1213 */
1214 task_ratelimit = (u64)dirty_ratelimit *
Tejun Heodaddfa32015-05-22 18:23:26 -04001215 dtc->pos_ratio >> RATELIMIT_CALC_SHIFT;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001216 task_ratelimit++; /* it helps rampup dirty_ratelimit from tiny values */
1217
1218 /*
1219 * A linear estimation of the "balanced" throttle rate. The theory is,
Tejun Heode1fff32015-05-22 17:13:29 -04001220 * if there are N dd tasks, each throttled at task_ratelimit, the wb's
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001221 * dirty_rate will be measured to be (N * task_ratelimit). So the below
1222 * formula will yield the balanced rate limit (write_bw / N).
1223 *
1224 * Note that the expanded form is not a pure rate feedback:
1225 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) (1)
1226 * but also takes pos_ratio into account:
1227 * rate_(i+1) = rate_(i) * (write_bw / dirty_rate) * pos_ratio (2)
1228 *
1229 * (1) is not realistic because pos_ratio also takes part in balancing
1230 * the dirty rate. Consider the state
1231 * pos_ratio = 0.5 (3)
1232 * rate = 2 * (write_bw / N) (4)
1233 * If (1) is used, it will stuck in that state! Because each dd will
1234 * be throttled at
1235 * task_ratelimit = pos_ratio * rate = (write_bw / N) (5)
1236 * yielding
1237 * dirty_rate = N * task_ratelimit = write_bw (6)
1238 * put (6) into (1) we get
1239 * rate_(i+1) = rate_(i) (7)
1240 *
1241 * So we end up using (2) to always keep
1242 * rate_(i+1) ~= (write_bw / N) (8)
1243 * regardless of the value of pos_ratio. As long as (8) is satisfied,
1244 * pos_ratio is able to drive itself to 1.0, which is not only where
1245 * the dirty count meet the setpoint, but also where the slope of
1246 * pos_ratio is most flat and hence task_ratelimit is least fluctuated.
1247 */
1248 balanced_dirty_ratelimit = div_u64((u64)task_ratelimit * write_bw,
1249 dirty_rate | 1);
Wu Fengguangbdaac492011-08-03 14:30:36 -06001250 /*
1251 * balanced_dirty_ratelimit ~= (write_bw / N) <= write_bw
1252 */
1253 if (unlikely(balanced_dirty_ratelimit > write_bw))
1254 balanced_dirty_ratelimit = write_bw;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001255
Wu Fengguang73811312011-08-26 15:53:24 -06001256 /*
1257 * We could safely do this and return immediately:
1258 *
Tejun Heode1fff32015-05-22 17:13:29 -04001259 * wb->dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguang73811312011-08-26 15:53:24 -06001260 *
1261 * However to get a more stable dirty_ratelimit, the below elaborated
Wanpeng Li331cbde2012-06-09 11:10:55 +08001262 * code makes use of task_ratelimit to filter out singular points and
Wu Fengguang73811312011-08-26 15:53:24 -06001263 * limit the step size.
1264 *
1265 * The below code essentially only uses the relative value of
1266 *
1267 * task_ratelimit - dirty_ratelimit
1268 * = (pos_ratio - 1) * dirty_ratelimit
1269 *
1270 * which reflects the direction and size of dirty position error.
1271 */
1272
1273 /*
1274 * dirty_ratelimit will follow balanced_dirty_ratelimit iff
1275 * task_ratelimit is on the same side of dirty_ratelimit, too.
1276 * For example, when
1277 * - dirty_ratelimit > balanced_dirty_ratelimit
1278 * - dirty_ratelimit > task_ratelimit (dirty pages are above setpoint)
1279 * lowering dirty_ratelimit will help meet both the position and rate
1280 * control targets. Otherwise, don't update dirty_ratelimit if it will
1281 * only help meet the rate target. After all, what the users ultimately
1282 * feel and care are stable dirty rate and small position error.
1283 *
1284 * |task_ratelimit - dirty_ratelimit| is used to limit the step size
Wanpeng Li331cbde2012-06-09 11:10:55 +08001285 * and filter out the singular points of balanced_dirty_ratelimit. Which
Wu Fengguang73811312011-08-26 15:53:24 -06001286 * keeps jumping around randomly and can even leap far away at times
1287 * due to the small 200ms estimation period of dirty_rate (we want to
1288 * keep that period small to reduce time lags).
1289 */
1290 step = 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001291
1292 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001293 * For strictlimit case, calculations above were based on wb counters
Tejun Heoa88a3412015-05-22 17:13:28 -04001294 * and limits (starting from pos_ratio = wb_position_ratio() and up to
Maxim Patlasov5a537482013-09-11 14:22:46 -07001295 * balanced_dirty_ratelimit = task_ratelimit * write_bw / dirty_rate).
Tejun Heode1fff32015-05-22 17:13:29 -04001296 * Hence, to calculate "step" properly, we have to use wb_dirty as
1297 * "dirty" and wb_setpoint as "setpoint".
Maxim Patlasov5a537482013-09-11 14:22:46 -07001298 *
Tejun Heode1fff32015-05-22 17:13:29 -04001299 * We rampup dirty_ratelimit forcibly if wb_dirty is low because
1300 * it's possible that wb_thresh is close to zero due to inactivity
Tejun Heo970fb012015-05-22 18:23:24 -04001301 * of backing device.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001302 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001303 if (unlikely(wb->bdi->capabilities & BDI_CAP_STRICTLIMIT)) {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001304 dirty = dtc->wb_dirty;
1305 if (dtc->wb_dirty < 8)
1306 setpoint = dtc->wb_dirty + 1;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001307 else
Tejun Heo970fb012015-05-22 18:23:24 -04001308 setpoint = (dtc->wb_thresh + dtc->wb_bg_thresh) / 2;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001309 }
1310
Wu Fengguang73811312011-08-26 15:53:24 -06001311 if (dirty < setpoint) {
Tejun Heoa88a3412015-05-22 17:13:28 -04001312 x = min3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001313 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001314 if (dirty_ratelimit < x)
1315 step = x - dirty_ratelimit;
1316 } else {
Tejun Heoa88a3412015-05-22 17:13:28 -04001317 x = max3(wb->balanced_dirty_ratelimit,
Mark Rustad7c809962014-10-09 15:28:15 -07001318 balanced_dirty_ratelimit, task_ratelimit);
Wu Fengguang73811312011-08-26 15:53:24 -06001319 if (dirty_ratelimit > x)
1320 step = dirty_ratelimit - x;
1321 }
1322
1323 /*
1324 * Don't pursue 100% rate matching. It's impossible since the balanced
1325 * rate itself is constantly fluctuating. So decrease the track speed
1326 * when it gets close to the target. Helps eliminate pointless tremors.
1327 */
Andrey Ryabinind59b1082016-03-15 14:55:27 -07001328 shift = dirty_ratelimit / (2 * step + 1);
1329 if (shift < BITS_PER_LONG)
1330 step = DIV_ROUND_UP(step >> shift, 8);
1331 else
1332 step = 0;
Wu Fengguang73811312011-08-26 15:53:24 -06001333
1334 if (dirty_ratelimit < balanced_dirty_ratelimit)
1335 dirty_ratelimit += step;
1336 else
1337 dirty_ratelimit -= step;
1338
Tejun Heoa88a3412015-05-22 17:13:28 -04001339 wb->dirty_ratelimit = max(dirty_ratelimit, 1UL);
1340 wb->balanced_dirty_ratelimit = balanced_dirty_ratelimit;
Wu Fengguangb48c1042011-03-02 17:22:49 -06001341
Tejun Heo5634cc22015-08-18 14:54:56 -07001342 trace_bdi_dirty_ratelimit(wb, dirty_rate, task_ratelimit);
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001343}
1344
Tejun Heoc2aa7232015-05-22 18:23:35 -04001345static void __wb_update_bandwidth(struct dirty_throttle_control *gdtc,
1346 struct dirty_throttle_control *mdtc,
Tejun Heo8a731792015-05-22 18:23:20 -04001347 unsigned long start_time,
1348 bool update_ratelimit)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001349{
Tejun Heoc2aa7232015-05-22 18:23:35 -04001350 struct bdi_writeback *wb = gdtc->wb;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001351 unsigned long now = jiffies;
Tejun Heoa88a3412015-05-22 17:13:28 -04001352 unsigned long elapsed = now - wb->bw_time_stamp;
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001353 unsigned long dirtied;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001354 unsigned long written;
1355
Tejun Heo8a731792015-05-22 18:23:20 -04001356 lockdep_assert_held(&wb->list_lock);
1357
Wu Fengguange98be2d2010-08-29 11:22:30 -06001358 /*
1359 * rate-limit, only update once every 200ms.
1360 */
1361 if (elapsed < BANDWIDTH_INTERVAL)
1362 return;
1363
Tejun Heoa88a3412015-05-22 17:13:28 -04001364 dirtied = percpu_counter_read(&wb->stat[WB_DIRTIED]);
1365 written = percpu_counter_read(&wb->stat[WB_WRITTEN]);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001366
1367 /*
1368 * Skip quiet periods when disk bandwidth is under-utilized.
1369 * (at least 1s idle time between two flusher runs)
1370 */
Tejun Heoa88a3412015-05-22 17:13:28 -04001371 if (elapsed > HZ && time_before(wb->bw_time_stamp, start_time))
Wu Fengguange98be2d2010-08-29 11:22:30 -06001372 goto snapshot;
1373
Tejun Heo8a731792015-05-22 18:23:20 -04001374 if (update_ratelimit) {
Tejun Heoc2aa7232015-05-22 18:23:35 -04001375 domain_update_bandwidth(gdtc, now);
1376 wb_update_dirty_ratelimit(gdtc, dirtied, elapsed);
1377
1378 /*
1379 * @mdtc is always NULL if !CGROUP_WRITEBACK but the
1380 * compiler has no way to figure that out. Help it.
1381 */
1382 if (IS_ENABLED(CONFIG_CGROUP_WRITEBACK) && mdtc) {
1383 domain_update_bandwidth(mdtc, now);
1384 wb_update_dirty_ratelimit(mdtc, dirtied, elapsed);
1385 }
Wu Fengguangbe3ffa22011-06-12 10:51:31 -06001386 }
Tejun Heoa88a3412015-05-22 17:13:28 -04001387 wb_update_write_bandwidth(wb, elapsed, written);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001388
1389snapshot:
Tejun Heoa88a3412015-05-22 17:13:28 -04001390 wb->dirtied_stamp = dirtied;
1391 wb->written_stamp = written;
1392 wb->bw_time_stamp = now;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001393}
1394
Tejun Heo8a731792015-05-22 18:23:20 -04001395void wb_update_bandwidth(struct bdi_writeback *wb, unsigned long start_time)
Wu Fengguange98be2d2010-08-29 11:22:30 -06001396{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001397 struct dirty_throttle_control gdtc = { GDTC_INIT(wb) };
1398
Tejun Heoc2aa7232015-05-22 18:23:35 -04001399 __wb_update_bandwidth(&gdtc, NULL, start_time, false);
Wu Fengguange98be2d2010-08-29 11:22:30 -06001400}
1401
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402/*
Namjae Jeond0e1d662012-12-11 16:00:21 -08001403 * After a task dirtied this many pages, balance_dirty_pages_ratelimited()
Wu Fengguang9d823e82011-06-11 18:10:12 -06001404 * will look to see if it needs to start dirty throttling.
1405 *
1406 * If dirty_poll_interval is too low, big NUMA machines will call the expensive
Michal Hockoc41f0122017-09-06 16:23:36 -07001407 * global_zone_page_state() too often. So scale it near-sqrt to the safety margin
Wu Fengguang9d823e82011-06-11 18:10:12 -06001408 * (the number of pages we may dirty without exceeding the dirty limits).
1409 */
1410static unsigned long dirty_poll_interval(unsigned long dirty,
1411 unsigned long thresh)
1412{
1413 if (thresh > dirty)
1414 return 1UL << (ilog2(thresh - dirty) >> 1);
1415
1416 return 1;
1417}
1418
Tejun Heoa88a3412015-05-22 17:13:28 -04001419static unsigned long wb_max_pause(struct bdi_writeback *wb,
Tejun Heode1fff32015-05-22 17:13:29 -04001420 unsigned long wb_dirty)
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001421{
Tejun Heoa88a3412015-05-22 17:13:28 -04001422 unsigned long bw = wb->avg_write_bandwidth;
Fengguang Wue3b6c652013-10-16 13:47:03 -07001423 unsigned long t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001424
1425 /*
1426 * Limit pause time for small memory systems. If sleeping for too long
1427 * time, a small pool of dirty/writeback pages may go empty and disk go
1428 * idle.
1429 *
1430 * 8 serves as the safety ratio.
1431 */
Tejun Heode1fff32015-05-22 17:13:29 -04001432 t = wb_dirty / (1 + bw / roundup_pow_of_two(1 + HZ / 8));
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001433 t++;
1434
Fengguang Wue3b6c652013-10-16 13:47:03 -07001435 return min_t(unsigned long, t, MAX_PAUSE);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001436}
1437
Tejun Heoa88a3412015-05-22 17:13:28 -04001438static long wb_min_pause(struct bdi_writeback *wb,
1439 long max_pause,
1440 unsigned long task_ratelimit,
1441 unsigned long dirty_ratelimit,
1442 int *nr_dirtied_pause)
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001443{
Tejun Heoa88a3412015-05-22 17:13:28 -04001444 long hi = ilog2(wb->avg_write_bandwidth);
1445 long lo = ilog2(wb->dirty_ratelimit);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001446 long t; /* target pause */
1447 long pause; /* estimated next pause */
1448 int pages; /* target nr_dirtied_pause */
1449
1450 /* target for 10ms pause on 1-dd case */
1451 t = max(1, HZ / 100);
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001452
1453 /*
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001454 * Scale up pause time for concurrent dirtiers in order to reduce CPU
1455 * overheads.
1456 *
1457 * (N * 10ms) on 2^N concurrent tasks.
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001458 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001459 if (hi > lo)
1460 t += (hi - lo) * (10 * HZ) / 1024;
1461
1462 /*
1463 * This is a bit convoluted. We try to base the next nr_dirtied_pause
1464 * on the much more stable dirty_ratelimit. However the next pause time
1465 * will be computed based on task_ratelimit and the two rate limits may
1466 * depart considerably at some time. Especially if task_ratelimit goes
1467 * below dirty_ratelimit/2 and the target pause is max_pause, the next
1468 * pause time will be max_pause*2 _trimmed down_ to max_pause. As a
1469 * result task_ratelimit won't be executed faithfully, which could
1470 * eventually bring down dirty_ratelimit.
1471 *
1472 * We apply two rules to fix it up:
1473 * 1) try to estimate the next pause time and if necessary, use a lower
1474 * nr_dirtied_pause so as not to exceed max_pause. When this happens,
1475 * nr_dirtied_pause will be "dancing" with task_ratelimit.
1476 * 2) limit the target pause time to max_pause/2, so that the normal
1477 * small fluctuations of task_ratelimit won't trigger rule (1) and
1478 * nr_dirtied_pause will remain as stable as dirty_ratelimit.
1479 */
1480 t = min(t, 1 + max_pause / 2);
1481 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1482
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001483 /*
1484 * Tiny nr_dirtied_pause is found to hurt I/O performance in the test
1485 * case fio-mmap-randwrite-64k, which does 16*{sync read, async write}.
1486 * When the 16 consecutive reads are often interrupted by some dirty
1487 * throttling pause during the async writes, cfq will go into idles
1488 * (deadline is fine). So push nr_dirtied_pause as high as possible
1489 * until reaches DIRTY_POLL_THRESH=32 pages.
1490 */
1491 if (pages < DIRTY_POLL_THRESH) {
1492 t = max_pause;
1493 pages = dirty_ratelimit * t / roundup_pow_of_two(HZ);
1494 if (pages > DIRTY_POLL_THRESH) {
1495 pages = DIRTY_POLL_THRESH;
1496 t = HZ * DIRTY_POLL_THRESH / dirty_ratelimit;
1497 }
1498 }
1499
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001500 pause = HZ * pages / (task_ratelimit + 1);
1501 if (pause > max_pause) {
1502 t = max_pause;
1503 pages = task_ratelimit * t / roundup_pow_of_two(HZ);
1504 }
1505
1506 *nr_dirtied_pause = pages;
1507 /*
1508 * The minimal pause time will normally be half the target pause time.
1509 */
Wu Fengguang5b9b3572011-12-06 13:17:17 -06001510 return pages >= DIRTY_POLL_THRESH ? 1 + t / 2 : t;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001511}
1512
Tejun Heo970fb012015-05-22 18:23:24 -04001513static inline void wb_dirty_limits(struct dirty_throttle_control *dtc)
Maxim Patlasov5a537482013-09-11 14:22:46 -07001514{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001515 struct bdi_writeback *wb = dtc->wb;
Tejun Heo93f78d82015-05-22 17:13:27 -04001516 unsigned long wb_reclaimable;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001517
1518 /*
Tejun Heode1fff32015-05-22 17:13:29 -04001519 * wb_thresh is not treated as some limiting factor as
Maxim Patlasov5a537482013-09-11 14:22:46 -07001520 * dirty_thresh, due to reasons
Tejun Heode1fff32015-05-22 17:13:29 -04001521 * - in JBOD setup, wb_thresh can fluctuate a lot
Maxim Patlasov5a537482013-09-11 14:22:46 -07001522 * - in a system with HDD and USB key, the USB key may somehow
Tejun Heode1fff32015-05-22 17:13:29 -04001523 * go into state (wb_dirty >> wb_thresh) either because
1524 * wb_dirty starts high, or because wb_thresh drops low.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001525 * In this case we don't want to hard throttle the USB key
Tejun Heode1fff32015-05-22 17:13:29 -04001526 * dirtiers for 100 seconds until wb_dirty drops under
1527 * wb_thresh. Instead the auxiliary wb control line in
Tejun Heoa88a3412015-05-22 17:13:28 -04001528 * wb_position_ratio() will let the dirtier task progress
Tejun Heode1fff32015-05-22 17:13:29 -04001529 * at some rate <= (write_bw / 2) for bringing down wb_dirty.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001530 */
Tejun Heob1cbc6d2015-05-22 18:23:25 -04001531 dtc->wb_thresh = __wb_calc_thresh(dtc);
Tejun Heo970fb012015-05-22 18:23:24 -04001532 dtc->wb_bg_thresh = dtc->thresh ?
1533 div_u64((u64)dtc->wb_thresh * dtc->bg_thresh, dtc->thresh) : 0;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001534
1535 /*
1536 * In order to avoid the stacked BDI deadlock we need
1537 * to ensure we accurately count the 'dirty' pages when
1538 * the threshold is low.
1539 *
1540 * Otherwise it would be possible to get thresh+n pages
1541 * reported dirty, even though there are thresh-m pages
1542 * actually dirty; with m+n sitting in the percpu
1543 * deltas.
1544 */
Wang Long2bce7742017-11-15 17:39:03 -08001545 if (dtc->wb_thresh < 2 * wb_stat_error()) {
Tejun Heo93f78d82015-05-22 17:13:27 -04001546 wb_reclaimable = wb_stat_sum(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001547 dtc->wb_dirty = wb_reclaimable + wb_stat_sum(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001548 } else {
Tejun Heo93f78d82015-05-22 17:13:27 -04001549 wb_reclaimable = wb_stat(wb, WB_RECLAIMABLE);
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001550 dtc->wb_dirty = wb_reclaimable + wb_stat(wb, WB_WRITEBACK);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001551 }
1552}
1553
Wu Fengguang9d823e82011-06-11 18:10:12 -06001554/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 * balance_dirty_pages() must be called by processes which are generating dirty
1556 * data. It looks at the number of dirty pages in the machine and will force
Wu Fengguang143dfe82010-08-27 18:45:12 -06001557 * the caller to wait once crossing the (background_thresh + dirty_thresh) / 2.
Jens Axboe5b0830c2009-09-23 19:37:09 +02001558 * If we're over `background_thresh' then the writeback threads are woken to
1559 * perform some writeout.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560 */
Tahsin Erdogan4c578dc2017-11-15 17:33:19 -08001561static void balance_dirty_pages(struct bdi_writeback *wb,
Wu Fengguang143dfe82010-08-27 18:45:12 -06001562 unsigned long pages_dirtied)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563{
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001564 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001565 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001566 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001567 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1568 &mdtc_stor : NULL;
1569 struct dirty_throttle_control *sdtc;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001570 unsigned long nr_reclaimable; /* = file_dirty + unstable_nfs */
Wu Fengguang83712352011-06-11 19:25:42 -06001571 long period;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001572 long pause;
1573 long max_pause;
1574 long min_pause;
1575 int nr_dirtied_pause;
Wu Fengguange50e3722010-08-11 14:17:37 -07001576 bool dirty_exceeded = false;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001577 unsigned long task_ratelimit;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001578 unsigned long dirty_ratelimit;
Tejun Heodfb8ae52015-05-22 17:13:40 -04001579 struct backing_dev_info *bdi = wb->bdi;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001580 bool strictlimit = bdi->capabilities & BDI_CAP_STRICTLIMIT;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001581 unsigned long start_time = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001582
1583 for (;;) {
Wu Fengguang83712352011-06-11 19:25:42 -06001584 unsigned long now = jiffies;
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001585 unsigned long dirty, thresh, bg_thresh;
Yang Shi50e55bf2015-11-20 15:57:10 -08001586 unsigned long m_dirty = 0; /* stop bogus uninit warnings */
1587 unsigned long m_thresh = 0;
1588 unsigned long m_bg_thresh = 0;
Wu Fengguang83712352011-06-11 19:25:42 -06001589
Wu Fengguang143dfe82010-08-27 18:45:12 -06001590 /*
1591 * Unstable writes are a feature of certain networked
1592 * filesystems (i.e. NFS) in which data may have been
1593 * written to the server's write cache, but has not yet
1594 * been flushed to permanent storage.
1595 */
Mel Gorman11fb9982016-07-28 15:46:20 -07001596 nr_reclaimable = global_node_page_state(NR_FILE_DIRTY) +
1597 global_node_page_state(NR_UNSTABLE_NFS);
Tejun Heo9fc3a432015-05-22 18:23:30 -04001598 gdtc->avail = global_dirtyable_memory();
Mel Gorman11fb9982016-07-28 15:46:20 -07001599 gdtc->dirty = nr_reclaimable + global_node_page_state(NR_WRITEBACK);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001600
Tejun Heo9fc3a432015-05-22 18:23:30 -04001601 domain_dirty_limits(gdtc);
Wu Fengguang16c40422010-08-11 14:17:39 -07001602
Maxim Patlasov5a537482013-09-11 14:22:46 -07001603 if (unlikely(strictlimit)) {
Tejun Heo970fb012015-05-22 18:23:24 -04001604 wb_dirty_limits(gdtc);
Maxim Patlasov5a537482013-09-11 14:22:46 -07001605
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001606 dirty = gdtc->wb_dirty;
1607 thresh = gdtc->wb_thresh;
Tejun Heo970fb012015-05-22 18:23:24 -04001608 bg_thresh = gdtc->wb_bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001609 } else {
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001610 dirty = gdtc->dirty;
1611 thresh = gdtc->thresh;
1612 bg_thresh = gdtc->bg_thresh;
Maxim Patlasov5a537482013-09-11 14:22:46 -07001613 }
1614
Tejun Heoc2aa7232015-05-22 18:23:35 -04001615 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001616 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001617
1618 /*
1619 * If @wb belongs to !root memcg, repeat the same
1620 * basic calculations for the memcg domain.
1621 */
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001622 mem_cgroup_wb_stats(wb, &filepages, &headroom,
1623 &mdtc->dirty, &writeback);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001624 mdtc->dirty += writeback;
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001625 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001626
1627 domain_dirty_limits(mdtc);
1628
1629 if (unlikely(strictlimit)) {
1630 wb_dirty_limits(mdtc);
1631 m_dirty = mdtc->wb_dirty;
1632 m_thresh = mdtc->wb_thresh;
1633 m_bg_thresh = mdtc->wb_bg_thresh;
1634 } else {
1635 m_dirty = mdtc->dirty;
1636 m_thresh = mdtc->thresh;
1637 m_bg_thresh = mdtc->bg_thresh;
1638 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001639 }
1640
1641 /*
1642 * Throttle it only when the background writeback cannot
1643 * catch-up. This avoids (excessively) small writeouts
Tejun Heode1fff32015-05-22 17:13:29 -04001644 * when the wb limits are ramping up in case of !strictlimit.
Maxim Patlasov5a537482013-09-11 14:22:46 -07001645 *
Tejun Heode1fff32015-05-22 17:13:29 -04001646 * In strictlimit case make decision based on the wb counters
1647 * and limits. Small writeouts when the wb limits are ramping
Maxim Patlasov5a537482013-09-11 14:22:46 -07001648 * up are the price we consciously pay for strictlimit-ing.
Tejun Heoc2aa7232015-05-22 18:23:35 -04001649 *
1650 * If memcg domain is in effect, @dirty should be under
1651 * both global and memcg freerun ceilings.
Wu Fengguang16c40422010-08-11 14:17:39 -07001652 */
Tejun Heoc2aa7232015-05-22 18:23:35 -04001653 if (dirty <= dirty_freerun_ceiling(thresh, bg_thresh) &&
1654 (!mdtc ||
1655 m_dirty <= dirty_freerun_ceiling(m_thresh, m_bg_thresh))) {
1656 unsigned long intv = dirty_poll_interval(dirty, thresh);
1657 unsigned long m_intv = ULONG_MAX;
1658
Wu Fengguang83712352011-06-11 19:25:42 -06001659 current->dirty_paused_when = now;
1660 current->nr_dirtied = 0;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001661 if (mdtc)
1662 m_intv = dirty_poll_interval(m_dirty, m_thresh);
1663 current->nr_dirtied_pause = min(intv, m_intv);
Wu Fengguang16c40422010-08-11 14:17:39 -07001664 break;
Wu Fengguang83712352011-06-11 19:25:42 -06001665 }
Wu Fengguang16c40422010-08-11 14:17:39 -07001666
Tejun Heobc058732015-05-22 17:13:53 -04001667 if (unlikely(!writeback_in_progress(wb)))
Tejun Heo9ecf48662015-05-22 17:13:54 -04001668 wb_start_background_writeback(wb);
Wu Fengguang143dfe82010-08-27 18:45:12 -06001669
Tejun Heoc2aa7232015-05-22 18:23:35 -04001670 /*
1671 * Calculate global domain's pos_ratio and select the
1672 * global dtc by default.
1673 */
Maxim Patlasov5a537482013-09-11 14:22:46 -07001674 if (!strictlimit)
Tejun Heo970fb012015-05-22 18:23:24 -04001675 wb_dirty_limits(gdtc);
Peter Zijlstra5fce25a2007-11-14 16:59:15 -08001676
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001677 dirty_exceeded = (gdtc->wb_dirty > gdtc->wb_thresh) &&
1678 ((gdtc->dirty > gdtc->thresh) || strictlimit);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679
Tejun Heodaddfa32015-05-22 18:23:26 -04001680 wb_position_ratio(gdtc);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001681 sdtc = gdtc;
Wu Fengguange98be2d2010-08-29 11:22:30 -06001682
Tejun Heoc2aa7232015-05-22 18:23:35 -04001683 if (mdtc) {
1684 /*
1685 * If memcg domain is in effect, calculate its
1686 * pos_ratio. @wb should satisfy constraints from
1687 * both global and memcg domains. Choose the one
1688 * w/ lower pos_ratio.
1689 */
1690 if (!strictlimit)
1691 wb_dirty_limits(mdtc);
1692
1693 dirty_exceeded |= (mdtc->wb_dirty > mdtc->wb_thresh) &&
1694 ((mdtc->dirty > mdtc->thresh) || strictlimit);
1695
1696 wb_position_ratio(mdtc);
1697 if (mdtc->pos_ratio < gdtc->pos_ratio)
1698 sdtc = mdtc;
1699 }
Tejun Heodaddfa32015-05-22 18:23:26 -04001700
Tejun Heoa88a3412015-05-22 17:13:28 -04001701 if (dirty_exceeded && !wb->dirty_exceeded)
1702 wb->dirty_exceeded = 1;
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07001703
Tejun Heo8a731792015-05-22 18:23:20 -04001704 if (time_is_before_jiffies(wb->bw_time_stamp +
1705 BANDWIDTH_INTERVAL)) {
1706 spin_lock(&wb->list_lock);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001707 __wb_update_bandwidth(gdtc, mdtc, start_time, true);
Tejun Heo8a731792015-05-22 18:23:20 -04001708 spin_unlock(&wb->list_lock);
1709 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710
Tejun Heoc2aa7232015-05-22 18:23:35 -04001711 /* throttle according to the chosen dtc */
Tejun Heoa88a3412015-05-22 17:13:28 -04001712 dirty_ratelimit = wb->dirty_ratelimit;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001713 task_ratelimit = ((u64)dirty_ratelimit * sdtc->pos_ratio) >>
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001714 RATELIMIT_CALC_SHIFT;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001715 max_pause = wb_max_pause(wb, sdtc->wb_dirty);
Tejun Heoa88a3412015-05-22 17:13:28 -04001716 min_pause = wb_min_pause(wb, max_pause,
1717 task_ratelimit, dirty_ratelimit,
1718 &nr_dirtied_pause);
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001719
Wu Fengguang3a73dbb2011-11-07 19:19:28 +08001720 if (unlikely(task_ratelimit == 0)) {
Wu Fengguang83712352011-06-11 19:25:42 -06001721 period = max_pause;
Wu Fengguangc8462cc2011-06-11 19:21:43 -06001722 pause = max_pause;
Wu Fengguang143dfe82010-08-27 18:45:12 -06001723 goto pause;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 }
Wu Fengguang83712352011-06-11 19:25:42 -06001725 period = HZ * pages_dirtied / task_ratelimit;
1726 pause = period;
1727 if (current->dirty_paused_when)
1728 pause -= now - current->dirty_paused_when;
1729 /*
1730 * For less than 1s think time (ext3/4 may block the dirtier
1731 * for up to 800ms from time to time on 1-HDD; so does xfs,
1732 * however at much less frequency), try to compensate it in
1733 * future periods by updating the virtual time; otherwise just
1734 * do a reset, as it may be a light dirtier.
1735 */
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001736 if (pause < min_pause) {
Tejun Heo5634cc22015-08-18 14:54:56 -07001737 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001738 sdtc->thresh,
1739 sdtc->bg_thresh,
1740 sdtc->dirty,
1741 sdtc->wb_thresh,
1742 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001743 dirty_ratelimit,
1744 task_ratelimit,
1745 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001746 period,
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001747 min(pause, 0L),
Wu Fengguangece13ac2010-08-29 23:33:20 -06001748 start_time);
Wu Fengguang83712352011-06-11 19:25:42 -06001749 if (pause < -HZ) {
1750 current->dirty_paused_when = now;
1751 current->nr_dirtied = 0;
1752 } else if (period) {
1753 current->dirty_paused_when += period;
1754 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001755 } else if (current->nr_dirtied_pause <= pages_dirtied)
1756 current->nr_dirtied_pause += pages_dirtied;
Wu Fengguang57fc9782011-06-11 19:32:32 -06001757 break;
1758 }
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001759 if (unlikely(pause > max_pause)) {
1760 /* for occasional dropped task_ratelimit */
1761 now += min(pause - max_pause, max_pause);
1762 pause = max_pause;
1763 }
Wu Fengguang143dfe82010-08-27 18:45:12 -06001764
1765pause:
Tejun Heo5634cc22015-08-18 14:54:56 -07001766 trace_balance_dirty_pages(wb,
Tejun Heoc2aa7232015-05-22 18:23:35 -04001767 sdtc->thresh,
1768 sdtc->bg_thresh,
1769 sdtc->dirty,
1770 sdtc->wb_thresh,
1771 sdtc->wb_dirty,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001772 dirty_ratelimit,
1773 task_ratelimit,
1774 pages_dirtied,
Wu Fengguang83712352011-06-11 19:25:42 -06001775 period,
Wu Fengguangece13ac2010-08-29 23:33:20 -06001776 pause,
1777 start_time);
Jan Kara499d05e2011-11-16 19:34:48 +08001778 __set_current_state(TASK_KILLABLE);
Jens Axboeb57d74a2016-09-01 10:20:33 -06001779 wb->dirty_sleep = now;
Wu Fengguangd25105e2009-10-09 12:40:42 +02001780 io_schedule_timeout(pause);
Jens Axboe87c6a9b2009-09-17 19:59:14 +02001781
Wu Fengguang83712352011-06-11 19:25:42 -06001782 current->dirty_paused_when = now + pause;
1783 current->nr_dirtied = 0;
Wu Fengguang7ccb9ad2011-11-30 11:08:55 -06001784 current->nr_dirtied_pause = nr_dirtied_pause;
Wu Fengguang83712352011-06-11 19:25:42 -06001785
Wu Fengguangffd1f602011-06-19 22:18:42 -06001786 /*
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001787 * This is typically equal to (dirty < thresh) and can also
1788 * keep "1000+ dd on a slow USB stick" under control.
Wu Fengguangffd1f602011-06-19 22:18:42 -06001789 */
Wu Fengguang1df64712011-11-13 19:47:32 -06001790 if (task_ratelimit)
Wu Fengguangffd1f602011-06-19 22:18:42 -06001791 break;
Jan Kara499d05e2011-11-16 19:34:48 +08001792
Wu Fengguangc5c63432011-12-02 10:21:33 -06001793 /*
1794 * In the case of an unresponding NFS server and the NFS dirty
Tejun Heode1fff32015-05-22 17:13:29 -04001795 * pages exceeds dirty_thresh, give the other good wb's a pipe
Wu Fengguangc5c63432011-12-02 10:21:33 -06001796 * to go through, so that tasks on them still remain responsive.
1797 *
Masahiro Yamada3f8b6fb2017-02-27 14:29:25 -08001798 * In theory 1 page is enough to keep the consumer-producer
Wu Fengguangc5c63432011-12-02 10:21:33 -06001799 * pipe going: the flusher cleans 1 page => the task dirties 1
Tejun Heode1fff32015-05-22 17:13:29 -04001800 * more page. However wb_dirty has accounting errors. So use
Tejun Heo93f78d82015-05-22 17:13:27 -04001801 * the larger and more IO friendly wb_stat_error.
Wu Fengguangc5c63432011-12-02 10:21:33 -06001802 */
Wang Long2bce7742017-11-15 17:39:03 -08001803 if (sdtc->wb_dirty <= wb_stat_error())
Wu Fengguangc5c63432011-12-02 10:21:33 -06001804 break;
1805
Jan Kara499d05e2011-11-16 19:34:48 +08001806 if (fatal_signal_pending(current))
1807 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 }
1809
Tejun Heoa88a3412015-05-22 17:13:28 -04001810 if (!dirty_exceeded && wb->dirty_exceeded)
1811 wb->dirty_exceeded = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812
Tejun Heobc058732015-05-22 17:13:53 -04001813 if (writeback_in_progress(wb))
Jens Axboe5b0830c2009-09-23 19:37:09 +02001814 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815
1816 /*
1817 * In laptop mode, we wait until hitting the higher threshold before
1818 * starting background writeout, and then write out all the way down
1819 * to the lower threshold. So slow writers cause minimal disk activity.
1820 *
1821 * In normal mode, we start background writeout at the lower
1822 * background_thresh, to keep the amount of dirty memory low.
1823 */
Wu Fengguang143dfe82010-08-27 18:45:12 -06001824 if (laptop_mode)
1825 return;
1826
Tejun Heo2bc00ae2015-05-22 18:23:23 -04001827 if (nr_reclaimable > gdtc->bg_thresh)
Tejun Heo9ecf48662015-05-22 17:13:54 -04001828 wb_start_background_writeback(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829}
1830
Wu Fengguang9d823e82011-06-11 18:10:12 -06001831static DEFINE_PER_CPU(int, bdp_ratelimits);
Tejun Heo245b2e72009-06-24 15:13:48 +09001832
Wu Fengguang54848d72011-04-05 13:21:19 -06001833/*
1834 * Normal tasks are throttled by
1835 * loop {
1836 * dirty tsk->nr_dirtied_pause pages;
1837 * take a snap in balance_dirty_pages();
1838 * }
1839 * However there is a worst case. If every task exit immediately when dirtied
1840 * (tsk->nr_dirtied_pause - 1) pages, balance_dirty_pages() will never be
1841 * called to throttle the page dirties. The solution is to save the not yet
1842 * throttled page dirties in dirty_throttle_leaks on task exit and charge them
1843 * randomly into the running tasks. This works well for the above worst case,
1844 * as the new task will pick up and accumulate the old task's leaked dirty
1845 * count and eventually get throttled.
1846 */
1847DEFINE_PER_CPU(int, dirty_throttle_leaks) = 0;
1848
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849/**
Namjae Jeond0e1d662012-12-11 16:00:21 -08001850 * balance_dirty_pages_ratelimited - balance dirty memory state
Martin Waitz67be2dd2005-05-01 08:59:26 -07001851 * @mapping: address_space which was dirtied
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 *
1853 * Processes which are dirtying memory should call in here once for each page
1854 * which was newly dirtied. The function will periodically check the system's
1855 * dirty state and will initiate writeback if needed.
1856 *
1857 * On really big machines, get_writeback_state is expensive, so try to avoid
1858 * calling it too often (ratelimiting). But once we're over the dirty memory
1859 * limit we decrease the ratelimiting by a lot, to prevent individual processes
1860 * from overshooting the limit by (ratelimit_pages) each.
1861 */
Namjae Jeond0e1d662012-12-11 16:00:21 -08001862void balance_dirty_pages_ratelimited(struct address_space *mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863{
Tejun Heodfb8ae52015-05-22 17:13:40 -04001864 struct inode *inode = mapping->host;
1865 struct backing_dev_info *bdi = inode_to_bdi(inode);
1866 struct bdi_writeback *wb = NULL;
Wu Fengguang9d823e82011-06-11 18:10:12 -06001867 int ratelimit;
1868 int *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869
Wu Fengguang36715ce2011-06-11 17:53:57 -06001870 if (!bdi_cap_account_dirty(bdi))
1871 return;
1872
Tejun Heodfb8ae52015-05-22 17:13:40 -04001873 if (inode_cgwb_enabled(inode))
1874 wb = wb_get_create_current(bdi, GFP_KERNEL);
1875 if (!wb)
1876 wb = &bdi->wb;
1877
Wu Fengguang9d823e82011-06-11 18:10:12 -06001878 ratelimit = current->nr_dirtied_pause;
Tejun Heoa88a3412015-05-22 17:13:28 -04001879 if (wb->dirty_exceeded)
Wu Fengguang9d823e82011-06-11 18:10:12 -06001880 ratelimit = min(ratelimit, 32 >> (PAGE_SHIFT - 10));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001882 preempt_disable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001883 /*
1884 * This prevents one CPU to accumulate too many dirtied pages without
1885 * calling into balance_dirty_pages(), which can happen when there are
1886 * 1000+ tasks, all of them start dirtying pages at exactly the same
1887 * time, hence all honoured too large initial task->nr_dirtied_pause.
1888 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001889 p = this_cpu_ptr(&bdp_ratelimits);
Wu Fengguang9d823e82011-06-11 18:10:12 -06001890 if (unlikely(current->nr_dirtied >= ratelimit))
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001891 *p = 0;
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06001892 else if (unlikely(*p >= ratelimit_pages)) {
1893 *p = 0;
1894 ratelimit = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 }
Wu Fengguang54848d72011-04-05 13:21:19 -06001896 /*
1897 * Pick up the dirtied pages by the exited tasks. This avoids lots of
1898 * short-lived tasks (eg. gcc invocations in a kernel build) escaping
1899 * the dirty throttling and livelock other long-run dirtiers.
1900 */
Christoph Lameter7c8e0182014-06-04 16:07:56 -07001901 p = this_cpu_ptr(&dirty_throttle_leaks);
Wu Fengguang54848d72011-04-05 13:21:19 -06001902 if (*p > 0 && current->nr_dirtied < ratelimit) {
Namjae Jeond0e1d662012-12-11 16:00:21 -08001903 unsigned long nr_pages_dirtied;
Wu Fengguang54848d72011-04-05 13:21:19 -06001904 nr_pages_dirtied = min(*p, ratelimit - current->nr_dirtied);
1905 *p -= nr_pages_dirtied;
1906 current->nr_dirtied += nr_pages_dirtied;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 }
Andrew Mortonfa5a7342006-03-24 03:18:10 -08001908 preempt_enable();
Wu Fengguang9d823e82011-06-11 18:10:12 -06001909
1910 if (unlikely(current->nr_dirtied >= ratelimit))
Tahsin Erdogan4c578dc2017-11-15 17:33:19 -08001911 balance_dirty_pages(wb, current->nr_dirtied);
Tejun Heodfb8ae52015-05-22 17:13:40 -04001912
1913 wb_put(wb);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914}
Namjae Jeond0e1d662012-12-11 16:00:21 -08001915EXPORT_SYMBOL(balance_dirty_pages_ratelimited);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
Tejun Heoaa661bb2015-05-22 18:23:31 -04001917/**
1918 * wb_over_bg_thresh - does @wb need to be written back?
1919 * @wb: bdi_writeback of interest
1920 *
1921 * Determines whether background writeback should keep writing @wb or it's
1922 * clean enough. Returns %true if writeback should continue.
1923 */
1924bool wb_over_bg_thresh(struct bdi_writeback *wb)
1925{
Tejun Heo947e9762015-05-22 18:23:32 -04001926 struct dirty_throttle_control gdtc_stor = { GDTC_INIT(wb) };
Tejun Heoc2aa7232015-05-22 18:23:35 -04001927 struct dirty_throttle_control mdtc_stor = { MDTC_INIT(wb, &gdtc_stor) };
Tejun Heo947e9762015-05-22 18:23:32 -04001928 struct dirty_throttle_control * const gdtc = &gdtc_stor;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001929 struct dirty_throttle_control * const mdtc = mdtc_valid(&mdtc_stor) ?
1930 &mdtc_stor : NULL;
Tejun Heoaa661bb2015-05-22 18:23:31 -04001931
Tejun Heo947e9762015-05-22 18:23:32 -04001932 /*
1933 * Similar to balance_dirty_pages() but ignores pages being written
1934 * as we're trying to decide whether to put more under writeback.
1935 */
1936 gdtc->avail = global_dirtyable_memory();
Mel Gorman11fb9982016-07-28 15:46:20 -07001937 gdtc->dirty = global_node_page_state(NR_FILE_DIRTY) +
1938 global_node_page_state(NR_UNSTABLE_NFS);
Tejun Heo947e9762015-05-22 18:23:32 -04001939 domain_dirty_limits(gdtc);
Tejun Heoaa661bb2015-05-22 18:23:31 -04001940
Tejun Heo947e9762015-05-22 18:23:32 -04001941 if (gdtc->dirty > gdtc->bg_thresh)
Tejun Heoaa661bb2015-05-22 18:23:31 -04001942 return true;
1943
Howard Cochran74d36942016-03-10 01:12:39 -05001944 if (wb_stat(wb, WB_RECLAIMABLE) >
1945 wb_calc_thresh(gdtc->wb, gdtc->bg_thresh))
Tejun Heoaa661bb2015-05-22 18:23:31 -04001946 return true;
1947
Tejun Heoc2aa7232015-05-22 18:23:35 -04001948 if (mdtc) {
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001949 unsigned long filepages, headroom, writeback;
Tejun Heoc2aa7232015-05-22 18:23:35 -04001950
Tejun Heoc5edf9c2015-09-29 13:04:26 -04001951 mem_cgroup_wb_stats(wb, &filepages, &headroom, &mdtc->dirty,
1952 &writeback);
1953 mdtc_calc_avail(mdtc, filepages, headroom);
Tejun Heoc2aa7232015-05-22 18:23:35 -04001954 domain_dirty_limits(mdtc); /* ditto, ignore writeback */
1955
1956 if (mdtc->dirty > mdtc->bg_thresh)
1957 return true;
1958
Howard Cochran74d36942016-03-10 01:12:39 -05001959 if (wb_stat(wb, WB_RECLAIMABLE) >
1960 wb_calc_thresh(mdtc->wb, mdtc->bg_thresh))
Tejun Heoc2aa7232015-05-22 18:23:35 -04001961 return true;
1962 }
1963
Tejun Heoaa661bb2015-05-22 18:23:31 -04001964 return false;
1965}
1966
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968 * sysctl handler for /proc/sys/vm/dirty_writeback_centisecs
1969 */
Joe Perchescccad5b2014-06-06 14:38:09 -07001970int dirty_writeback_centisecs_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07001971 void __user *buffer, size_t *length, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972{
Yafang Shao94af5842017-10-10 05:44:13 +08001973 unsigned int old_interval = dirty_writeback_interval;
1974 int ret;
1975
1976 ret = proc_dointvec(table, write, buffer, length, ppos);
Yafang Shao515c24c2017-10-14 16:38:27 +08001977
1978 /*
1979 * Writing 0 to dirty_writeback_interval will disable periodic writeback
1980 * and a different non-zero value will wakeup the writeback threads.
1981 * wb_wakeup_delayed() would be more appropriate, but it's a pain to
1982 * iterate over all bdis and wbs.
1983 * The reason we do this is to make the change take effect immediately.
1984 */
1985 if (!ret && write && dirty_writeback_interval &&
1986 dirty_writeback_interval != old_interval)
Yafang Shao94af5842017-10-10 05:44:13 +08001987 wakeup_flusher_threads(WB_REASON_PERIODIC);
1988
1989 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990}
1991
Jens Axboec2c49862010-05-20 09:18:47 +02001992#ifdef CONFIG_BLOCK
Kees Cookbca237a2017-08-28 15:03:41 -07001993void laptop_mode_timer_fn(struct timer_list *t)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994{
Kees Cookbca237a2017-08-28 15:03:41 -07001995 struct backing_dev_info *backing_dev_info =
1996 from_timer(backing_dev_info, t, laptop_mode_wb_timer);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997
Kees Cookbca237a2017-08-28 15:03:41 -07001998 wakeup_flusher_threads_bdi(backing_dev_info, WB_REASON_LAPTOP_TIMER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999}
2000
2001/*
2002 * We've spun up the disk and we're in laptop mode: schedule writeback
2003 * of all dirty data a few seconds from now. If the flush is already scheduled
2004 * then push it back - the user is still using the disk.
2005 */
Matthew Garrett31373d02010-04-06 14:25:14 +02002006void laptop_io_completion(struct backing_dev_info *info)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007{
Matthew Garrett31373d02010-04-06 14:25:14 +02002008 mod_timer(&info->laptop_mode_wb_timer, jiffies + laptop_mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009}
2010
2011/*
2012 * We're in laptop mode and we've just synced. The sync's writes will have
2013 * caused another writeback to be scheduled by laptop_io_completion.
2014 * Nothing needs to be written back anymore, so we unschedule the writeback.
2015 */
2016void laptop_sync_completion(void)
2017{
Matthew Garrett31373d02010-04-06 14:25:14 +02002018 struct backing_dev_info *bdi;
2019
2020 rcu_read_lock();
2021
2022 list_for_each_entry_rcu(bdi, &bdi_list, bdi_list)
2023 del_timer(&bdi->laptop_mode_wb_timer);
2024
2025 rcu_read_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002026}
Jens Axboec2c49862010-05-20 09:18:47 +02002027#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
2029/*
2030 * If ratelimit_pages is too high then we can get into dirty-data overload
2031 * if a large number of processes all perform writes at the same time.
2032 * If it is too low then SMP machines will call the (expensive)
2033 * get_writeback_state too often.
2034 *
2035 * Here we set ratelimit_pages to a level which ensures that when all CPUs are
2036 * dirtying in parallel, we cannot go more than 3% (1/32) over the dirty memory
Wu Fengguang9d823e82011-06-11 18:10:12 -06002037 * thresholds.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 */
2039
Chandra Seetharaman2d1d43f2006-09-29 02:01:25 -07002040void writeback_set_ratelimit(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041{
Tejun Heodcc25ae2015-05-22 18:23:22 -04002042 struct wb_domain *dom = &global_wb_domain;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002043 unsigned long background_thresh;
2044 unsigned long dirty_thresh;
Tejun Heodcc25ae2015-05-22 18:23:22 -04002045
Wu Fengguang9d823e82011-06-11 18:10:12 -06002046 global_dirty_limits(&background_thresh, &dirty_thresh);
Tejun Heodcc25ae2015-05-22 18:23:22 -04002047 dom->dirty_limit = dirty_thresh;
Wu Fengguang9d823e82011-06-11 18:10:12 -06002048 ratelimit_pages = dirty_thresh / (num_online_cpus() * 32);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049 if (ratelimit_pages < 16)
2050 ratelimit_pages = 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051}
2052
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002053static int page_writeback_cpu_online(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054{
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002055 writeback_set_ratelimit();
2056 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057}
2058
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059/*
Linus Torvaldsdc6e29d2007-01-29 16:37:38 -08002060 * Called early on to tune the page writeback dirty limits.
2061 *
2062 * We used to scale dirty pages according to how total memory
2063 * related to pages that could be allocated for buffers (by
2064 * comparing nr_free_buffer_pages() to vm_total_pages.
2065 *
2066 * However, that was when we used "dirty_ratio" to scale with
2067 * all memory, and we don't do that any more. "dirty_ratio"
2068 * is now applied to total non-HIGHPAGE memory (by subtracting
2069 * totalhigh_pages from vm_total_pages), and as such we can't
2070 * get into the old insane situation any more where we had
2071 * large amounts of dirty pages compared to a small amount of
2072 * non-HIGHMEM memory.
2073 *
2074 * But we might still want to scale the dirty_ratio by how
2075 * much memory the box has..
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 */
2077void __init page_writeback_init(void)
2078{
Rabin Vincenta50fcb52015-08-06 15:47:14 -07002079 BUG_ON(wb_domain_init(&global_wb_domain, GFP_KERNEL));
2080
Sebastian Andrzej Siewior1d7ac6a2016-08-18 14:57:20 +02002081 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "mm/writeback:online",
2082 page_writeback_cpu_online, NULL);
2083 cpuhp_setup_state(CPUHP_MM_WRITEBACK_DEAD, "mm/writeback:dead", NULL,
2084 page_writeback_cpu_online);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085}
2086
David Howells811d7362006-08-29 19:06:09 +01002087/**
Jan Karaf446daae2010-08-09 17:19:12 -07002088 * tag_pages_for_writeback - tag pages to be written by write_cache_pages
2089 * @mapping: address space structure to write
2090 * @start: starting page index
2091 * @end: ending page index (inclusive)
2092 *
2093 * This function scans the page range from @start to @end (inclusive) and tags
2094 * all pages that have DIRTY tag set with a special TOWRITE tag. The idea is
2095 * that write_cache_pages (or whoever calls this function) will then use
2096 * TOWRITE tag to identify pages eligible for writeback. This mechanism is
2097 * used to avoid livelocking of writeback by a process steadily creating new
2098 * dirty pages in the file (thus it is important for this function to be quick
2099 * so that it can tag pages faster than a dirtying process can create them).
2100 */
2101/*
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002102 * We tag pages in batches of WRITEBACK_TAG_BATCH to reduce the i_pages lock
2103 * latency.
Jan Karaf446daae2010-08-09 17:19:12 -07002104 */
Jan Karaf446daae2010-08-09 17:19:12 -07002105void tag_pages_for_writeback(struct address_space *mapping,
2106 pgoff_t start, pgoff_t end)
2107{
Randy Dunlap3c111a02010-08-11 14:17:30 -07002108#define WRITEBACK_TAG_BATCH 4096
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002109 unsigned long tagged = 0;
2110 struct radix_tree_iter iter;
2111 void **slot;
Jan Karaf446daae2010-08-09 17:19:12 -07002112
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002113 xa_lock_irq(&mapping->i_pages);
2114 radix_tree_for_each_tagged(slot, &mapping->i_pages, &iter, start,
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002115 PAGECACHE_TAG_DIRTY) {
2116 if (iter.index > end)
2117 break;
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002118 radix_tree_iter_tag_set(&mapping->i_pages, &iter,
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002119 PAGECACHE_TAG_TOWRITE);
2120 tagged++;
2121 if ((tagged % WRITEBACK_TAG_BATCH) != 0)
2122 continue;
2123 slot = radix_tree_iter_resume(slot, &iter);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002124 xa_unlock_irq(&mapping->i_pages);
Jan Karaf446daae2010-08-09 17:19:12 -07002125 cond_resched();
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002126 xa_lock_irq(&mapping->i_pages);
Matthew Wilcox268f42d2016-12-14 15:08:55 -08002127 }
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002128 xa_unlock_irq(&mapping->i_pages);
Jan Karaf446daae2010-08-09 17:19:12 -07002129}
2130EXPORT_SYMBOL(tag_pages_for_writeback);
2131
2132/**
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002133 * write_cache_pages - walk the list of dirty pages of the given address space and write all of them.
David Howells811d7362006-08-29 19:06:09 +01002134 * @mapping: address space structure to write
2135 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002136 * @writepage: function called for each page
2137 * @data: data passed to writepage function
David Howells811d7362006-08-29 19:06:09 +01002138 *
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002139 * If a page is already under I/O, write_cache_pages() skips it, even
David Howells811d7362006-08-29 19:06:09 +01002140 * if it's dirty. This is desirable behaviour for memory-cleaning writeback,
2141 * but it is INCORRECT for data-integrity system calls such as fsync(). fsync()
2142 * and msync() need to guarantee that all the data which was dirty at the time
2143 * the call was made get new I/O started against them. If wbc->sync_mode is
2144 * WB_SYNC_ALL then we were called for data integrity and we must wait for
2145 * existing IO to complete.
Jan Karaf446daae2010-08-09 17:19:12 -07002146 *
2147 * To avoid livelocks (when other process dirties new pages), we first tag
2148 * pages which should be written back with TOWRITE tag and only then start
2149 * writing them. For data-integrity sync we have to be careful so that we do
2150 * not miss some pages (e.g., because some other process has cleared TOWRITE
2151 * tag we set). The rule we follow is that TOWRITE tag can be cleared only
2152 * by the process clearing the DIRTY tag (and submitting the page for IO).
David Howells811d7362006-08-29 19:06:09 +01002153 */
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002154int write_cache_pages(struct address_space *mapping,
2155 struct writeback_control *wbc, writepage_t writepage,
2156 void *data)
David Howells811d7362006-08-29 19:06:09 +01002157{
David Howells811d7362006-08-29 19:06:09 +01002158 int ret = 0;
2159 int done = 0;
David Howells811d7362006-08-29 19:06:09 +01002160 struct pagevec pvec;
2161 int nr_pages;
Nick Piggin31a12662009-01-06 14:39:04 -08002162 pgoff_t uninitialized_var(writeback_index);
David Howells811d7362006-08-29 19:06:09 +01002163 pgoff_t index;
2164 pgoff_t end; /* Inclusive */
Nick Pigginbd19e012009-01-06 14:39:06 -08002165 pgoff_t done_index;
Nick Piggin31a12662009-01-06 14:39:04 -08002166 int cycled;
David Howells811d7362006-08-29 19:06:09 +01002167 int range_whole = 0;
Jan Karaf446daae2010-08-09 17:19:12 -07002168 int tag;
David Howells811d7362006-08-29 19:06:09 +01002169
Mel Gorman86679822017-11-15 17:37:52 -08002170 pagevec_init(&pvec);
David Howells811d7362006-08-29 19:06:09 +01002171 if (wbc->range_cyclic) {
Nick Piggin31a12662009-01-06 14:39:04 -08002172 writeback_index = mapping->writeback_index; /* prev offset */
2173 index = writeback_index;
2174 if (index == 0)
2175 cycled = 1;
2176 else
2177 cycled = 0;
David Howells811d7362006-08-29 19:06:09 +01002178 end = -1;
2179 } else {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002180 index = wbc->range_start >> PAGE_SHIFT;
2181 end = wbc->range_end >> PAGE_SHIFT;
David Howells811d7362006-08-29 19:06:09 +01002182 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
2183 range_whole = 1;
Nick Piggin31a12662009-01-06 14:39:04 -08002184 cycled = 1; /* ignore range_cyclic tests */
David Howells811d7362006-08-29 19:06:09 +01002185 }
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002186 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002187 tag = PAGECACHE_TAG_TOWRITE;
2188 else
2189 tag = PAGECACHE_TAG_DIRTY;
David Howells811d7362006-08-29 19:06:09 +01002190retry:
Wu Fengguang6e6938b2010-06-06 10:38:15 -06002191 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
Jan Karaf446daae2010-08-09 17:19:12 -07002192 tag_pages_for_writeback(mapping, index, end);
Nick Pigginbd19e012009-01-06 14:39:06 -08002193 done_index = index;
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002194 while (!done && (index <= end)) {
2195 int i;
2196
Jan Kara2b9775a2017-11-15 17:35:09 -08002197 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
Jan Kara67fd7072017-11-15 17:35:19 -08002198 tag);
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002199 if (nr_pages == 0)
2200 break;
David Howells811d7362006-08-29 19:06:09 +01002201
David Howells811d7362006-08-29 19:06:09 +01002202 for (i = 0; i < nr_pages; i++) {
2203 struct page *page = pvec.pages[i];
2204
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002205 done_index = page->index;
Nick Pigginbd19e012009-01-06 14:39:06 -08002206
David Howells811d7362006-08-29 19:06:09 +01002207 lock_page(page);
2208
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002209 /*
2210 * Page truncated or invalidated. We can freely skip it
2211 * then, even for data integrity operations: the page
2212 * has disappeared concurrently, so there could be no
2213 * real expectation of this data interity operation
2214 * even if there is now a new, dirty page at the same
2215 * pagecache address.
2216 */
David Howells811d7362006-08-29 19:06:09 +01002217 if (unlikely(page->mapping != mapping)) {
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002218continue_unlock:
David Howells811d7362006-08-29 19:06:09 +01002219 unlock_page(page);
2220 continue;
2221 }
2222
Nick Piggin515f4a02009-01-06 14:39:10 -08002223 if (!PageDirty(page)) {
2224 /* someone wrote it for us */
2225 goto continue_unlock;
2226 }
David Howells811d7362006-08-29 19:06:09 +01002227
Nick Piggin515f4a02009-01-06 14:39:10 -08002228 if (PageWriteback(page)) {
2229 if (wbc->sync_mode != WB_SYNC_NONE)
2230 wait_on_page_writeback(page);
2231 else
2232 goto continue_unlock;
2233 }
2234
2235 BUG_ON(PageWriteback(page));
2236 if (!clear_page_dirty_for_io(page))
Nick Piggin5a3d5c92009-01-06 14:39:09 -08002237 goto continue_unlock;
David Howells811d7362006-08-29 19:06:09 +01002238
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002239 trace_wbc_writepage(wbc, inode_to_bdi(mapping->host));
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002240 ret = (*writepage)(page, wbc, data);
Nick Piggin00266772009-01-06 14:39:06 -08002241 if (unlikely(ret)) {
2242 if (ret == AOP_WRITEPAGE_ACTIVATE) {
2243 unlock_page(page);
2244 ret = 0;
2245 } else {
2246 /*
2247 * done_index is set past this page,
2248 * so media errors will not choke
2249 * background writeout for the entire
2250 * file. This has consequences for
2251 * range_cyclic semantics (ie. it may
2252 * not be suitable for data integrity
2253 * writeout).
2254 */
Jun'ichi Nomuracf15b072011-03-22 16:33:40 -07002255 done_index = page->index + 1;
Nick Piggin00266772009-01-06 14:39:06 -08002256 done = 1;
2257 break;
2258 }
Dave Chinner0b564922010-06-09 10:37:18 +10002259 }
David Howells811d7362006-08-29 19:06:09 +01002260
Dave Chinner546a1922010-08-24 11:44:34 +10002261 /*
2262 * We stop writing back only if we are not doing
2263 * integrity sync. In case of integrity sync we have to
2264 * keep going until we have written all the pages
2265 * we tagged for writeback prior to entering this loop.
2266 */
2267 if (--wbc->nr_to_write <= 0 &&
2268 wbc->sync_mode == WB_SYNC_NONE) {
2269 done = 1;
2270 break;
Nick Piggin05fe4782009-01-06 14:39:08 -08002271 }
David Howells811d7362006-08-29 19:06:09 +01002272 }
2273 pagevec_release(&pvec);
2274 cond_resched();
2275 }
Nick Piggin3a4c6802009-02-12 04:34:23 +01002276 if (!cycled && !done) {
David Howells811d7362006-08-29 19:06:09 +01002277 /*
Nick Piggin31a12662009-01-06 14:39:04 -08002278 * range_cyclic:
David Howells811d7362006-08-29 19:06:09 +01002279 * We hit the last page and there is more work to be done: wrap
2280 * back to the start of the file
2281 */
Nick Piggin31a12662009-01-06 14:39:04 -08002282 cycled = 1;
David Howells811d7362006-08-29 19:06:09 +01002283 index = 0;
Nick Piggin31a12662009-01-06 14:39:04 -08002284 end = writeback_index - 1;
David Howells811d7362006-08-29 19:06:09 +01002285 goto retry;
2286 }
Dave Chinner0b564922010-06-09 10:37:18 +10002287 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
2288 mapping->writeback_index = done_index;
Aneesh Kumar K.V06d6cf62008-07-11 19:27:31 -04002289
David Howells811d7362006-08-29 19:06:09 +01002290 return ret;
2291}
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002292EXPORT_SYMBOL(write_cache_pages);
2293
2294/*
2295 * Function used by generic_writepages to call the real writepage
2296 * function and set the mapping flags on error
2297 */
2298static int __writepage(struct page *page, struct writeback_control *wbc,
2299 void *data)
2300{
2301 struct address_space *mapping = data;
2302 int ret = mapping->a_ops->writepage(page, wbc);
2303 mapping_set_error(mapping, ret);
2304 return ret;
2305}
2306
2307/**
2308 * generic_writepages - walk the list of dirty pages of the given address space and writepage() all of them.
2309 * @mapping: address space structure to write
2310 * @wbc: subtract the number of written pages from *@wbc->nr_to_write
2311 *
2312 * This is a library function, which implements the writepages()
2313 * address_space_operation.
2314 */
2315int generic_writepages(struct address_space *mapping,
2316 struct writeback_control *wbc)
2317{
Shaohua Li9b6096a2011-03-17 10:47:06 +01002318 struct blk_plug plug;
2319 int ret;
2320
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002321 /* deal with chardevs and other special file */
2322 if (!mapping->a_ops->writepage)
2323 return 0;
2324
Shaohua Li9b6096a2011-03-17 10:47:06 +01002325 blk_start_plug(&plug);
2326 ret = write_cache_pages(mapping, wbc, __writepage, mapping);
2327 blk_finish_plug(&plug);
2328 return ret;
Miklos Szeredi0ea97182007-05-10 22:22:51 -07002329}
David Howells811d7362006-08-29 19:06:09 +01002330
2331EXPORT_SYMBOL(generic_writepages);
2332
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333int do_writepages(struct address_space *mapping, struct writeback_control *wbc)
2334{
Andrew Morton22905f72005-11-16 15:07:01 -08002335 int ret;
2336
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 if (wbc->nr_to_write <= 0)
2338 return 0;
Theodore Ts'o80a2ea92017-04-28 09:51:54 -04002339 while (1) {
2340 if (mapping->a_ops->writepages)
2341 ret = mapping->a_ops->writepages(mapping, wbc);
2342 else
2343 ret = generic_writepages(mapping, wbc);
2344 if ((ret != -ENOMEM) || (wbc->sync_mode != WB_SYNC_ALL))
2345 break;
2346 cond_resched();
2347 congestion_wait(BLK_RW_ASYNC, HZ/50);
2348 }
Andrew Morton22905f72005-11-16 15:07:01 -08002349 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350}
2351
2352/**
Jeff Layton2b69c822017-07-05 15:26:48 -04002353 * write_one_page - write out a single page and wait on I/O
Martin Waitz67be2dd2005-05-01 08:59:26 -07002354 * @page: the page to write
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 *
2356 * The page must be locked by the caller and will be unlocked upon return.
2357 *
Jeff Layton37e51a72017-07-05 15:17:19 -04002358 * Note that the mapping's AS_EIO/AS_ENOSPC flags will be cleared when this
2359 * function returns.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360 */
Jeff Layton2b69c822017-07-05 15:26:48 -04002361int write_one_page(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362{
2363 struct address_space *mapping = page->mapping;
2364 int ret = 0;
2365 struct writeback_control wbc = {
2366 .sync_mode = WB_SYNC_ALL,
2367 .nr_to_write = 1,
2368 };
2369
2370 BUG_ON(!PageLocked(page));
2371
Jeff Layton2b69c822017-07-05 15:26:48 -04002372 wait_on_page_writeback(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373
2374 if (clear_page_dirty_for_io(page)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002375 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376 ret = mapping->a_ops->writepage(page, &wbc);
Jeff Layton37e51a72017-07-05 15:17:19 -04002377 if (ret == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 wait_on_page_writeback(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002379 put_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 } else {
2381 unlock_page(page);
2382 }
Jeff Layton37e51a72017-07-05 15:17:19 -04002383
2384 if (!ret)
2385 ret = filemap_check_errors(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 return ret;
2387}
2388EXPORT_SYMBOL(write_one_page);
2389
2390/*
Ken Chen76719322007-02-10 01:43:15 -08002391 * For address_spaces which do not use buffers nor write back.
2392 */
2393int __set_page_dirty_no_writeback(struct page *page)
2394{
2395 if (!PageDirty(page))
Bob Liuc3f0da62011-01-13 15:45:49 -08002396 return !TestSetPageDirty(page);
Ken Chen76719322007-02-10 01:43:15 -08002397 return 0;
2398}
2399
2400/*
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002401 * Helper function for set_page_dirty family.
Greg Thelenc4843a72015-05-22 17:13:16 -04002402 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002403 * Caller must hold lock_page_memcg().
Greg Thelenc4843a72015-05-22 17:13:16 -04002404 *
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002405 * NOTE: This relies on being atomic wrt interrupts.
2406 */
Johannes Weiner62cccb82016-03-15 14:57:22 -07002407void account_page_dirtied(struct page *page, struct address_space *mapping)
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002408{
Tejun Heo52ebea72015-05-22 17:13:37 -04002409 struct inode *inode = mapping->host;
2410
Tejun Heo9fb0a7d2013-01-11 13:06:37 -08002411 trace_writeback_dirty_page(page, mapping);
2412
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002413 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo52ebea72015-05-22 17:13:37 -04002414 struct bdi_writeback *wb;
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002415
Tejun Heo52ebea72015-05-22 17:13:37 -04002416 inode_attach_wb(inode, page);
2417 wb = inode_to_wb(inode);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002418
Johannes Weiner00f3ca22017-07-06 15:40:52 -07002419 __inc_lruvec_page_state(page, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002420 __inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Mel Gormanc4a25632016-07-28 15:46:23 -07002421 __inc_node_page_state(page, NR_DIRTIED);
Nikolay Borisov3e8f3992017-07-12 14:37:51 -07002422 inc_wb_stat(wb, WB_RECLAIMABLE);
2423 inc_wb_stat(wb, WB_DIRTIED);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002424 task_io_account_write(PAGE_SIZE);
Wu Fengguangd3bc1fe2011-04-14 07:52:37 -06002425 current->nr_dirtied++;
2426 this_cpu_inc(bdp_ratelimits);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002427 }
2428}
Michael Rubin679ceac2010-08-20 02:31:26 -07002429EXPORT_SYMBOL(account_page_dirtied);
Edward Shishkine3a7cca2009-03-31 15:19:39 -07002430
2431/*
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002432 * Helper function for deaccounting dirty page without writeback.
2433 *
Johannes Weiner81f8c3a2016-03-15 14:57:04 -07002434 * Caller must hold lock_page_memcg().
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002435 */
Greg Thelenc4843a72015-05-22 17:13:16 -04002436void account_page_cleaned(struct page *page, struct address_space *mapping,
Johannes Weiner62cccb82016-03-15 14:57:22 -07002437 struct bdi_writeback *wb)
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002438{
2439 if (mapping_cap_account_dirty(mapping)) {
Johannes Weiner00f3ca22017-07-06 15:40:52 -07002440 dec_lruvec_page_state(page, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002441 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002442 dec_wb_stat(wb, WB_RECLAIMABLE);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002443 task_io_account_cancelled_write(PAGE_SIZE);
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002444 }
2445}
Konstantin Khlebnikovb9ea2512015-04-14 15:45:27 -07002446
2447/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 * For address_spaces which do not use buffers. Just tag the page as dirty in
2449 * its radix tree.
2450 *
2451 * This is also used when a single buffer is being dirtied: we want to set the
2452 * page dirty in that case, but not all the buffers. This is a "bottom-up"
2453 * dirtying, whereas __set_page_dirty_buffers() is a "top-down" dirtying.
2454 *
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002455 * The caller must ensure this doesn't race with truncation. Most will simply
2456 * hold the page lock, but e.g. zap_pte_range() calls with the page mapped and
2457 * the pte lock held, which also locks out truncation.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 */
2459int __set_page_dirty_nobuffers(struct page *page)
2460{
Johannes Weiner62cccb82016-03-15 14:57:22 -07002461 lock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 if (!TestSetPageDirty(page)) {
2463 struct address_space *mapping = page_mapping(page);
KOSAKI Motohiroa85d9df2014-02-06 12:04:24 -08002464 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465
Greg Thelenc4843a72015-05-22 17:13:16 -04002466 if (!mapping) {
Johannes Weiner62cccb82016-03-15 14:57:22 -07002467 unlock_page_memcg(page);
Andrew Morton8c085402006-12-10 02:19:24 -08002468 return 1;
Greg Thelenc4843a72015-05-22 17:13:16 -04002469 }
Andrew Morton8c085402006-12-10 02:19:24 -08002470
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002471 xa_lock_irqsave(&mapping->i_pages, flags);
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002472 BUG_ON(page_mapping(page) != mapping);
2473 WARN_ON_ONCE(!PagePrivate(page) && !PageUptodate(page));
Johannes Weiner62cccb82016-03-15 14:57:22 -07002474 account_page_dirtied(page, mapping);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002475 radix_tree_tag_set(&mapping->i_pages, page_index(page),
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002476 PAGECACHE_TAG_DIRTY);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002477 xa_unlock_irqrestore(&mapping->i_pages, flags);
Johannes Weiner62cccb82016-03-15 14:57:22 -07002478 unlock_page_memcg(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002479
Andrew Morton8c085402006-12-10 02:19:24 -08002480 if (mapping->host) {
2481 /* !PageAnon && !swapper_space */
2482 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002484 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07002486 unlock_page_memcg(page);
Andrew Morton4741c9f2006-03-24 03:18:11 -08002487 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488}
2489EXPORT_SYMBOL(__set_page_dirty_nobuffers);
2490
2491/*
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002492 * Call this whenever redirtying a page, to de-account the dirty counters
2493 * (NR_DIRTIED, BDI_DIRTIED, tsk->nr_dirtied), so that they match the written
2494 * counters (NR_WRITTEN, BDI_WRITTEN) in long term. The mismatches will lead to
2495 * systematic errors in balanced_dirty_ratelimit and the dirty pages position
2496 * control.
2497 */
2498void account_page_redirty(struct page *page)
2499{
2500 struct address_space *mapping = page->mapping;
Tejun Heo91018132015-05-22 17:13:39 -04002501
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002502 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002503 struct inode *inode = mapping->host;
2504 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002505 struct wb_lock_cookie cookie = {};
Tejun Heo91018132015-05-22 17:13:39 -04002506
Greg Thelen2e898e42018-04-20 14:55:42 -07002507 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002508 current->nr_dirtied--;
Mel Gormanc4a25632016-07-28 15:46:23 -07002509 dec_node_page_state(page, NR_DIRTIED);
Tejun Heo91018132015-05-22 17:13:39 -04002510 dec_wb_stat(wb, WB_DIRTIED);
Greg Thelen2e898e42018-04-20 14:55:42 -07002511 unlocked_inode_to_wb_end(inode, &cookie);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002512 }
2513}
2514EXPORT_SYMBOL(account_page_redirty);
2515
2516/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 * When a writepage implementation decides that it doesn't want to write this
2518 * page for some reason, it should redirty the locked page via
2519 * redirty_page_for_writepage() and it should then unlock the page and return 0
2520 */
2521int redirty_page_for_writepage(struct writeback_control *wbc, struct page *page)
2522{
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002523 int ret;
2524
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 wbc->pages_skipped++;
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002526 ret = __set_page_dirty_nobuffers(page);
Wu Fengguang2f800fb2011-08-08 15:22:00 -06002527 account_page_redirty(page);
Konstantin Khebnikov8d386332015-02-11 15:26:55 -08002528 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529}
2530EXPORT_SYMBOL(redirty_page_for_writepage);
2531
2532/*
Wu Fengguang6746aff2009-09-16 11:50:14 +02002533 * Dirty a page.
2534 *
2535 * For pages with a mapping this should be done under the page lock
2536 * for the benefit of asynchronous memory errors who prefer a consistent
2537 * dirty state. This rule can be broken in some special cases,
2538 * but should be better not to.
2539 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 * If the mapping doesn't provide a set_page_dirty a_op, then
2541 * just fall through and assume that it wants buffer_heads.
2542 */
Nick Piggin1cf6e7d2009-02-18 14:48:18 -08002543int set_page_dirty(struct page *page)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544{
2545 struct address_space *mapping = page_mapping(page);
2546
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07002547 page = compound_head(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548 if (likely(mapping)) {
2549 int (*spd)(struct page *) = mapping->a_ops->set_page_dirty;
Minchan Kim278df9f2011-03-22 16:32:54 -07002550 /*
2551 * readahead/lru_deactivate_page could remain
2552 * PG_readahead/PG_reclaim due to race with end_page_writeback
2553 * About readahead, if the page is written, the flags would be
2554 * reset. So no problem.
2555 * About lru_deactivate_page, if the page is redirty, the flag
2556 * will be reset. So no problem. but if the page is used by readahead
2557 * it will confuse readahead and make it restart the size rampup
2558 * process. But it's a trivial problem.
2559 */
Naoya Horiguchia4bb3ec2015-04-15 16:13:17 -07002560 if (PageReclaim(page))
2561 ClearPageReclaim(page);
David Howells93614012006-09-30 20:45:40 +02002562#ifdef CONFIG_BLOCK
2563 if (!spd)
2564 spd = __set_page_dirty_buffers;
2565#endif
2566 return (*spd)(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 }
Andrew Morton4741c9f2006-03-24 03:18:11 -08002568 if (!PageDirty(page)) {
2569 if (!TestSetPageDirty(page))
2570 return 1;
2571 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 return 0;
2573}
2574EXPORT_SYMBOL(set_page_dirty);
2575
2576/*
2577 * set_page_dirty() is racy if the caller has no reference against
2578 * page->mapping->host, and if the page is unlocked. This is because another
2579 * CPU could truncate the page off the mapping and then free the mapping.
2580 *
2581 * Usually, the page _is_ locked, or the caller is a user-space process which
2582 * holds a reference on the inode by having an open file.
2583 *
2584 * In other cases, the page should be locked before running set_page_dirty().
2585 */
2586int set_page_dirty_lock(struct page *page)
2587{
2588 int ret;
2589
Jens Axboe7eaceac2011-03-10 08:52:07 +01002590 lock_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 ret = set_page_dirty(page);
2592 unlock_page(page);
2593 return ret;
2594}
2595EXPORT_SYMBOL(set_page_dirty_lock);
2596
2597/*
Tejun Heo11f81be2015-05-22 17:13:15 -04002598 * This cancels just the dirty bit on the kernel page itself, it does NOT
2599 * actually remove dirty bits on any mmap's that may be around. It also
2600 * leaves the page tagged dirty, so any sync activity will still find it on
2601 * the dirty lists, and in particular, clear_page_dirty_for_io() will still
2602 * look at the dirty bits in the VM.
2603 *
2604 * Doing this should *normally* only ever be done when a page is truncated,
2605 * and is not actually mapped anywhere at all. However, fs/buffer.c does
2606 * this when it notices that somebody has cleaned out all the buffers on a
2607 * page without actually doing it through the VM. Can you say "ext3 is
2608 * horribly ugly"? Thought you could.
2609 */
Jan Kara736304f2017-11-15 17:37:11 -08002610void __cancel_dirty_page(struct page *page)
Tejun Heo11f81be2015-05-22 17:13:15 -04002611{
Greg Thelenc4843a72015-05-22 17:13:16 -04002612 struct address_space *mapping = page_mapping(page);
2613
2614 if (mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002615 struct inode *inode = mapping->host;
2616 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002617 struct wb_lock_cookie cookie = {};
Greg Thelenc4843a72015-05-22 17:13:16 -04002618
Johannes Weiner62cccb82016-03-15 14:57:22 -07002619 lock_page_memcg(page);
Greg Thelen2e898e42018-04-20 14:55:42 -07002620 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002621
2622 if (TestClearPageDirty(page))
Johannes Weiner62cccb82016-03-15 14:57:22 -07002623 account_page_cleaned(page, mapping, wb);
Greg Thelenc4843a72015-05-22 17:13:16 -04002624
Greg Thelen2e898e42018-04-20 14:55:42 -07002625 unlocked_inode_to_wb_end(inode, &cookie);
Johannes Weiner62cccb82016-03-15 14:57:22 -07002626 unlock_page_memcg(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002627 } else {
2628 ClearPageDirty(page);
2629 }
Tejun Heo11f81be2015-05-22 17:13:15 -04002630}
Jan Kara736304f2017-11-15 17:37:11 -08002631EXPORT_SYMBOL(__cancel_dirty_page);
Tejun Heo11f81be2015-05-22 17:13:15 -04002632
2633/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634 * Clear a page's dirty flag, while caring for dirty memory accounting.
2635 * Returns true if the page was previously dirty.
2636 *
2637 * This is for preparing to put the page under writeout. We leave the page
2638 * tagged as dirty in the radix tree so that a concurrent write-for-sync
2639 * can discover it via a PAGECACHE_TAG_DIRTY walk. The ->writepage
2640 * implementation will run either set_page_writeback() or set_page_dirty(),
2641 * at which stage we bring the page's dirty flag and radix-tree dirty tag
2642 * back into sync.
2643 *
2644 * This incoherency between the page's dirty flag and radix-tree tag is
2645 * unfortunate, but it only exists while the page is locked.
2646 */
2647int clear_page_dirty_for_io(struct page *page)
2648{
2649 struct address_space *mapping = page_mapping(page);
Greg Thelenc4843a72015-05-22 17:13:16 -04002650 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651
Nick Piggin79352892007-07-19 01:47:22 -07002652 BUG_ON(!PageLocked(page));
2653
Linus Torvalds7658cc22006-12-29 10:00:58 -08002654 if (mapping && mapping_cap_account_dirty(mapping)) {
Tejun Heo682aa8e2015-05-28 14:50:53 -04002655 struct inode *inode = mapping->host;
2656 struct bdi_writeback *wb;
Greg Thelen2e898e42018-04-20 14:55:42 -07002657 struct wb_lock_cookie cookie = {};
Tejun Heo682aa8e2015-05-28 14:50:53 -04002658
Linus Torvalds7658cc22006-12-29 10:00:58 -08002659 /*
2660 * Yes, Virginia, this is indeed insane.
2661 *
2662 * We use this sequence to make sure that
2663 * (a) we account for dirty stats properly
2664 * (b) we tell the low-level filesystem to
2665 * mark the whole page dirty if it was
2666 * dirty in a pagetable. Only to then
2667 * (c) clean the page again and return 1 to
2668 * cause the writeback.
2669 *
2670 * This way we avoid all nasty races with the
2671 * dirty bit in multiple places and clearing
2672 * them concurrently from different threads.
2673 *
2674 * Note! Normally the "set_page_dirty(page)"
2675 * has no effect on the actual dirty bit - since
2676 * that will already usually be set. But we
2677 * need the side effects, and it can help us
2678 * avoid races.
2679 *
2680 * We basically use the page "master dirty bit"
2681 * as a serialization point for all the different
2682 * threads doing their things.
Linus Torvalds7658cc22006-12-29 10:00:58 -08002683 */
2684 if (page_mkclean(page))
2685 set_page_dirty(page);
Nick Piggin79352892007-07-19 01:47:22 -07002686 /*
2687 * We carefully synchronise fault handlers against
2688 * installing a dirty pte and marking the page dirty
Johannes Weiner2d6d7f92015-01-08 14:32:18 -08002689 * at this point. We do this by having them hold the
2690 * page lock while dirtying the page, and pages are
2691 * always locked coming in here, so we get the desired
2692 * exclusion.
Nick Piggin79352892007-07-19 01:47:22 -07002693 */
Greg Thelen2e898e42018-04-20 14:55:42 -07002694 wb = unlocked_inode_to_wb_begin(inode, &cookie);
Linus Torvalds7658cc22006-12-29 10:00:58 -08002695 if (TestClearPageDirty(page)) {
Johannes Weiner00f3ca22017-07-06 15:40:52 -07002696 dec_lruvec_page_state(page, NR_FILE_DIRTY);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002697 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Tejun Heo682aa8e2015-05-28 14:50:53 -04002698 dec_wb_stat(wb, WB_RECLAIMABLE);
Greg Thelenc4843a72015-05-22 17:13:16 -04002699 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700 }
Greg Thelen2e898e42018-04-20 14:55:42 -07002701 unlocked_inode_to_wb_end(inode, &cookie);
Greg Thelenc4843a72015-05-22 17:13:16 -04002702 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 }
Linus Torvalds7658cc22006-12-29 10:00:58 -08002704 return TestClearPageDirty(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705}
Hans Reiser58bb01a2005-11-18 01:10:53 -08002706EXPORT_SYMBOL(clear_page_dirty_for_io);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707
2708int test_clear_page_writeback(struct page *page)
2709{
2710 struct address_space *mapping = page_mapping(page);
Johannes Weiner739f79f2017-08-18 15:15:48 -07002711 struct mem_cgroup *memcg;
2712 struct lruvec *lruvec;
Johannes Weinerd7365e72014-10-29 14:50:48 -07002713 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714
Johannes Weiner739f79f2017-08-18 15:15:48 -07002715 memcg = lock_page_memcg(page);
2716 lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));
Huang Ying371a0962016-10-07 16:59:30 -07002717 if (mapping && mapping_use_writeback_tags(mapping)) {
Tejun Heo91018132015-05-22 17:13:39 -04002718 struct inode *inode = mapping->host;
2719 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 unsigned long flags;
2721
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002722 xa_lock_irqsave(&mapping->i_pages, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 ret = TestClearPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002724 if (ret) {
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002725 radix_tree_tag_clear(&mapping->i_pages, page_index(page),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002727 if (bdi_cap_account_writeback(bdi)) {
Tejun Heo91018132015-05-22 17:13:39 -04002728 struct bdi_writeback *wb = inode_to_wb(inode);
2729
Nikolay Borisov3e8f3992017-07-12 14:37:51 -07002730 dec_wb_stat(wb, WB_WRITEBACK);
Tejun Heo91018132015-05-22 17:13:39 -04002731 __wb_writeout_inc(wb);
Peter Zijlstra04fbfdc2007-10-16 23:25:50 -07002732 }
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002733 }
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002734
2735 if (mapping->host && !mapping_tagged(mapping,
2736 PAGECACHE_TAG_WRITEBACK))
2737 sb_clear_inode_writeback(mapping->host);
2738
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002739 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 } else {
2741 ret = TestClearPageWriteback(page);
2742 }
Johannes Weiner739f79f2017-08-18 15:15:48 -07002743 /*
2744 * NOTE: Page might be free now! Writeback doesn't hold a page
2745 * reference on its own, it relies on truncation to wait for
2746 * the clearing of PG_writeback. The below can only access
2747 * page state that is static across allocation cycles.
2748 */
Wu Fengguang99b12e32011-07-25 17:12:37 -07002749 if (ret) {
Johannes Weiner739f79f2017-08-18 15:15:48 -07002750 dec_lruvec_state(lruvec, NR_WRITEBACK);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002751 dec_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Mel Gormanc4a25632016-07-28 15:46:23 -07002752 inc_node_page_state(page, NR_WRITTEN);
Wu Fengguang99b12e32011-07-25 17:12:37 -07002753 }
Johannes Weiner739f79f2017-08-18 15:15:48 -07002754 __unlock_page_memcg(memcg);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 return ret;
2756}
2757
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002758int __test_set_page_writeback(struct page *page, bool keep_write)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759{
2760 struct address_space *mapping = page_mapping(page);
Johannes Weinerd7365e72014-10-29 14:50:48 -07002761 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762
Johannes Weiner62cccb82016-03-15 14:57:22 -07002763 lock_page_memcg(page);
Huang Ying371a0962016-10-07 16:59:30 -07002764 if (mapping && mapping_use_writeback_tags(mapping)) {
Tejun Heo91018132015-05-22 17:13:39 -04002765 struct inode *inode = mapping->host;
2766 struct backing_dev_info *bdi = inode_to_bdi(inode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 unsigned long flags;
2768
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002769 xa_lock_irqsave(&mapping->i_pages, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 ret = TestSetPageWriteback(page);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002771 if (!ret) {
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002772 bool on_wblist;
2773
2774 on_wblist = mapping_tagged(mapping,
2775 PAGECACHE_TAG_WRITEBACK);
2776
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002777 radix_tree_tag_set(&mapping->i_pages, page_index(page),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 PAGECACHE_TAG_WRITEBACK);
Miklos Szeredie4ad08f2008-04-30 00:54:37 -07002779 if (bdi_cap_account_writeback(bdi))
Nikolay Borisov3e8f3992017-07-12 14:37:51 -07002780 inc_wb_stat(inode_to_wb(inode), WB_WRITEBACK);
Dave Chinner6c60d2b2016-07-26 15:21:50 -07002781
2782 /*
2783 * We can come through here when swapping anonymous
2784 * pages, so we don't necessarily have an inode to track
2785 * for sync.
2786 */
2787 if (mapping->host && !on_wblist)
2788 sb_mark_inode_writeback(mapping->host);
Peter Zijlstra69cb51d2007-10-16 23:25:48 -07002789 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 if (!PageDirty(page))
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002791 radix_tree_tag_clear(&mapping->i_pages, page_index(page),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 PAGECACHE_TAG_DIRTY);
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002793 if (!keep_write)
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002794 radix_tree_tag_clear(&mapping->i_pages, page_index(page),
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002795 PAGECACHE_TAG_TOWRITE);
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002796 xa_unlock_irqrestore(&mapping->i_pages, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 } else {
2798 ret = TestSetPageWriteback(page);
2799 }
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002800 if (!ret) {
Johannes Weiner00f3ca22017-07-06 15:40:52 -07002801 inc_lruvec_page_state(page, NR_WRITEBACK);
Mel Gorman5a1c84b2016-07-28 15:47:31 -07002802 inc_zone_page_state(page, NR_ZONE_WRITE_PENDING);
Johannes Weiner3a3c02e2014-10-29 14:50:46 -07002803 }
Johannes Weiner62cccb82016-03-15 14:57:22 -07002804 unlock_page_memcg(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 return ret;
2806
2807}
Namjae Jeon1c8349a2014-05-12 08:12:25 -04002808EXPORT_SYMBOL(__test_set_page_writeback);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809
2810/*
Nick Piggin00128182007-10-16 01:24:40 -07002811 * Return true if any of the pages in the mapping are marked with the
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 * passed tag.
2813 */
2814int mapping_tagged(struct address_space *mapping, int tag)
2815{
Matthew Wilcoxb93b0162018-04-10 16:36:56 -07002816 return radix_tree_tagged(&mapping->i_pages, tag);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817}
2818EXPORT_SYMBOL(mapping_tagged);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002819
2820/**
2821 * wait_for_stable_page() - wait for writeback to finish, if necessary.
2822 * @page: The page to wait on.
2823 *
2824 * This function determines if the given page is related to a backing device
2825 * that requires page contents to be held stable during writeback. If so, then
2826 * it will wait for any pending writeback to complete.
2827 */
2828void wait_for_stable_page(struct page *page)
2829{
Christoph Hellwigde1414a2015-01-14 10:42:36 +01002830 if (bdi_cap_stable_pages_required(inode_to_bdi(page->mapping->host)))
2831 wait_on_page_writeback(page);
Darrick J. Wong1d1d1a72013-02-21 16:42:51 -08002832}
2833EXPORT_SYMBOL_GPL(wait_for_stable_page);